research work on ergonomics

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Publications
Account settings
My Bibliography
Collections
Citation manager

Save citation to file

Email citation, add to collections.

Create a new collection
Add to an existing collection

Add to My Bibliography

Your saved search, create a file for external citation management software, your rss feed.

Search in PubMed
Search in NLM Catalog
Add to Search

Evidence-based ergonomics: a model and conceptual structure proposal

Affiliation.

1 Industrial Systems, Petroleum, Gas and Energy Laboratory, Industrial and Metallurgical Engineering School,Universidade Federal Fluminens, Av. dos Trabalhadores 420, Volta Redonda, RJ, Brazil. [email protected]
PMID: 22316771
DOI: 10.3233/WOR-2012-0201-487

In Human Factors and Ergonomics Science (HFES), it is difficult to identify what is the best approach to tackle the workplace and systems design problems which needs to be solved, and it has been also advocated as transdisciplinary and multidisciplinary the issue of "How to solve the human factors and ergonomics problems that are identified?". The proposition on this study is to combine the theoretical approach for Sustainability Science, the Taxonomy of the Human Factors and Ergonomics (HFE) discipline and the framework for Evidence-Based Medicine in an attempt to be applied in Human Factors and Ergonomics. Applications of ontologies are known in the field of medical research and computer science. By scrutinizing the key requirements for the HFES structuring of knowledge, it was designed a reference model, First, it was identified the important requirements for HFES Concept structuring, as regarded by Meister. Second, it was developed an evidence-based ergonomics framework as a reference model composed of six levels based on these requirements. Third, it was devised a mapping tool using linguistic resources to translate human work, systems environment and the complexities inherent to their hierarchical relationships to support future development at Level 2 of the reference model and for meeting the two major challenges for HFES, namely, identifying what problems should be addressed in HFE as an Autonomous Science itself and proposing solutions by integrating concepts and methods applied in HFES for those problems.

PubMed Disclaimer

LinkOut - more resources

Full text sources.

MedlinePlus Health Information
Citation Manager

NCBI Literature Resources

MeSH PMC Bookshelf Disclaimer

The PubMed wordmark and PubMed logo are registered trademarks of the U.S. Department of Health and Human Services (HHS). Unauthorized use of these marks is strictly prohibited.

REVIEW article

Neuroergonomics: a review of applications to physical and cognitive work.

$\r\nRanjana K. Mehta*$

1 Department of Environmental and Occupational Health, School of Rural Public Healthy, Texas A&M University, College Station, TX, USA
2 Center of Excellence in Neuroergonomics, Technology, and Cognition, George Mason University, Fairfax, VA, USA

Neuroergonomics is an emerging science that is defined as the study of the human brain in relation to performance at work and in everyday settings. This paper provides a critical review of the neuroergonomic approach to evaluating physical and cognitive work, particularly in mobile settings. Neuroergonomics research employing mobile and immobile brain imaging techniques are discussed in the following areas of physical and cognitive work: (1) physical work parameters; (2) physical fatigue; (3) vigilance and mental fatigue; (4) training and neuroadaptive systems; and (5) assessment of concurrent physical and cognitive work. Finally, the integration of brain and body measurements in investigating workload and fatigue, in the context of mobile brain/body imaging (“MoBI”), is discussed.

Introduction

Neuroergonomics is defined as the study of the human brain in relation to performance at work and everyday settings ( Parasuraman, 2003 ; Parasuraman and Rizzo, 2007 ). It integrates theories and principles from ergonomics, neuroscience, and human factors to provide valuable insights on brain function and behavior as encountered in natural settings ( Parasuraman, 2011 ). In this paper, we review neuroimaging techniques applicable to neuroergonomics that has expanded our understanding of the neural correlates of operators’ physical and cognitive capabilities and limitations when they interact with work systems. Moreover, while experimental laboratory studies have advanced our knowledge of brain functions during simulated work, it is important to assess operator performance in naturalistic work settings. Understanding brain function in such dynamic and mobile work settings requires the use of ambulatory neuroimaging techniques ( Makeig et al., 2009 ).

There are two main reasons why ambulatory neuroimaging techniques need to be developed for ergonomics research and practice. First, by definition, physical ergonomics requires that participants move their limbs or bodies while carrying out some physical task. Moreover, while cognitive ergonomics studies can be conducted in immobile participants, research on embodied cognition has shown that cognitive processing when moving and interacting in the physical world may have unique characteristics that can only be captured with mobile neuroimaging ( Clark, 1998 ; Parasuraman, 2003 ; Raz et al., 2005 ). This review discusses the use of neuroergonomics methods to evaluate brain responses in mobile work environments. We discuss the suitability and feasibility of mobile and immobile brain imaging techniques in the context of physical neuroergonomics, cognitive neuroergonomics, and neuroergonomic assessment of concurrent physical and mental work. Finally, we consider the requirements and utility of combined brain and body measurements in investigating workload and fatigue for neuroergonomic investigations.

Neuroergonomic Methods

Neuroergonomic studies rely heavily on existing neuroimaging techniques to understand brain structures, mechanisms, and functions during work. Neuroimaging techniques applicable to neuroergonomics fall into two general categories, those that are direct indicators of neuronal activity in response to stimuli, such as electroencephalography (EEG) and event-related potentials (ERPs), and those that provide indirect metabolic indicators of neuronal activity, such as functional magnetic resonance imaging (fMRI), positron emission tomography (PET), and functional near infrared spectroscopy (fNIRS). EEG represents summated post-synaptic electrical activity of neurons firing in response to motor/cognitive stimuli as recorded on the scalp, and thus offers excellent temporal resolution of electromagnetic brain changes, on the order of milliseconds. In comparison, fMRI and PET techniques, that provide information on cerebral blood flow in response to neuronal activity, have low temporal resolution (on the order of about 10 s), but offer excellent spatial resolution (1 cm or better) and unlike EEG, they provide valuable information on location of the neural signal generated.

Since neuroergonomics distinguishes itself from traditional neuroscience in that it evaluates brain functions in response to work, it is important that the neuroergonomic methods provide the flexibility to assess brain function in naturalistic work settings. Some neuroimaging techniques are better designed for and adapted for assessing brain functions in mobile work environments than others. The pros and cons of neuroergonomic methods are discussed in reference to three criteria: (1) temporal resolution, (2) spatial resolution, and (3) degree of immobility. Figure 1 illustrates how these neuroimaging techniques compare against each other based on the three criteria. In addition, Table 1 lists these methods and their major characteristics, such as portability, cost, along with spatial and temporal resolution. In this section, we provide a brief review of the various methods that have been used in neuroergonomic evaluations of human work, emphasizing measures of brain function and applicability in mobile experimental/field settings.

FIGURE 1. A comparison of electromagnetic (pink) and hemodynamic (blue) neuroimaging techniques for use in neuroergonomics based on temporal resolution ( x -axis), spatial resolution ( y -axis), and degree of immobility ( z -axis). EEG, electroencephalography; ERP, event-related potentials; MEG, magnetoencephalography; fNIRS, functional near infrared spectroscopy; TCDS, transcranial Doppler sonography; fMRI, functional magnetic resonance imaging; DTI, Diffusion tensor imaging; PET, positron emission tomography.

TABLE 1. List of neuroergonomic techniques and their major features.

Electroencephalography signals are the spatial summation of current density induced by synchronized post-synaptic potentials occurring in large clusters of neurons measured at the scalp ( Pizzagalli, 2007 ). The EEG is recorded as differences in voltage between active electrodes at different positions on the scalp, such as the frontal, parietal, temporal, and occipital lobes of the brain according to the International 10–20 System, and a reference electrode, typically the ear. EEG signals comprises of different frequency bands, each associated with various cognitive and physical states. Spectral analyses of EEG signals can be conducted to assess power in different frequency bands: delta (0.5–3 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (40–50 Hz). Another commonly computed EEG-driven spectral metric (i.e., brain) used in conjunction with muscular output (i.e., body) is corticomuscular coherence (CMC). CMC reflects “communications” between the brain and muscle and is determined as the coherence between sensorimotor cortex activation obtained from EEG and muscular activation as measured by electromyography (EMG) during motor activities ( Halliday et al., 1995 ).

Electroencephalography-derived ERPs represent the brain’s neural response to specific sensory, motor, and cognitive events. ERPs represent the outcome of signal averaging of EEG epochs time-locked to a particular stimulus or response event. To evaluate mental workload or examine human error ( Fedota and Parasuraman, 2010 ), ERP waveforms are examined for changes in the amplitude and latency of different ERP components , typically defined as positive or negative peak activity (such as the P3 and N1 components) or slowly rising activity such as the lateralized readiness potential ( Luck, 2005 ). To assess neural bases of motor activities, motor-related cortical potential (MRCP) ERP components have been studied that are characterized by a slowly rising negative potential, called Bereitschaftspotential (BP) or readiness potential, which is followed by a sharp rising negative potential, known as negative slope. As the onset of MRCP occurs prior to the onset of the motor activity, MRCP is considered to indicate pre-motor activity, which involves specific brain regions that prepare for a desired motor behavior ( Kornhuber and Deecke, 1965 ).

Electroencephalography-driven metrics, both spectral and temporal, in evaluating brain function during naturalistic complex tasks are relatively unobtrusive so that it does not interfere with the operator’s work performance. Its compact size and low cost, compared to other neuroimaging techniques such as fMRI and PET, makes it fairly well suited for use in both laboratory and field conditions. While artifacts attributed to movement, eye blinks, and physiological interference accompany EEG data, several algorithms have been developed to allow for the removal of noise in the EEG signal in real time or during post processing of the data ( Jung et al., 2000 ). Recent developments in making “field-friendly” EEG systems include “dry” electrode caps, which do not need extensive participant preparation time, as well as wireless systems that do not require the participant to be tethered to cables. These technical developments have enhanced the relevance and value of EEG for mobile applications ( Makeig et al., 2009 ).

Cerebral hemodynamic techniques such as fMRI and PET provide valuable information on source locations of distinct neural activation patterns associated with simple and complex cognitive, motor, and affective functions. While PET uses injected radioactive tracers to measure the blood flow in response to stimuli, based on their respective magnetic characteristics fMRI focuses on the resulting contrast between oxygenated and deoxygenated blood called the Blood Oxygenation Level Dependent or BOLD signal ( Poldrack et al., 2011 ). Both fMRI and PET have been fundamental in advancing our knowledge on brain functions and mechanisms during simple, and relatively static, cognitive and motor tasks. By leveraging high spatial resolution offered by fMRI measurements, reliable techniques for the fMRI-EEG integration have been made possible that offer greater spatio-temporal resolution of imaging dynamic brain activity as well as significant improvement over the conventional fMRI-weighted EEG source imaging techniques ( Liu and He, 2008 ; Yang et al., 2010 ). At the same time, fMRI and PET present several limitations in studying brain functions, such as the required supine position that may yield altered hemodynamic changes than seated or standing positions ( Raz et al., 2005 ), limited mobility, and restrictions on synchronized brain-body measurements ( Makeig et al., 2009 ). Moreover, the increasing need to examine brain activation patterns in complex tasks more representative of natural everyday situations have led researchers to adopt alternative neuroimaging techniques that offer better mobility features.

Functional near infrared spectroscopy is a non-invasive optical technique for measuring cerebral hemodynamics similar to PET and fMRI but with lower spatial resolution. By utilizing the tight neurovascular coupling between neuronal activity and regional cerebral blood flow ( Villringer and Chance, 1997 ) fNIRS measures regional cerebral hemodynamic changes (i.e., changes in oxy- and deoxy-hemoglobin levels) ( Jobsis, 1977 ). Since oxygenated and deoxygenated blood can be contrasted by their different optical absorption properties, fNIRS detects the levels of these blood parameters in response to neuronal activity. fNIRS is portable, inexpensive, and has shown to be an effective tool in quantifying cortical activation during static and dynamic motor movements, without causing substantial movement artifact issues ( Perrey, 2008 ). While fNIRS measurements, particularly oxygenated hemoglobin levels, have shown to be strongly correlated to the fMRI BOLD signals, albeit with relatively lower signal to noise ratio ( Strangman et al., 2002 ; Cui et al., 2011 ), unlike fMRI and PET its effectiveness in mapping neural activations across closely connected regions or within deep cortical areas is limited due to its relatively lower spatial resolution. Multimodal imaging approaches using both fNIRS and EEG systems have demonstrated that fNIRS is capable of enhancing event-related desynchronization-based EEG measurements significantly ( Leamy et al., 2011 ; Fazli et al., 2012 ).

While fNIRS enables measurement of oxygenated and deoxygenated hemoglobin levels in cortical regions, transcranial Doppler sonography (TCDS) uses ultrasound to image cerebral blood flow to the brain hemispheres ( Aaslid, 1986 ). TCDS uses an emitter attached to the head to direct ultrasound toward the middle cerebral artery (MCA) within the brain, and a receiver then records the frequency of the sound wave reflected by red blood cells moving through the artery. The magnitude of the change in frequency (the Doppler shift) varies directly proportional to the velocity of blood flow within the artery ( Duschek and Schandry, 2003 ). In response to increased task-related neuronal activity, MCA blood flow velocity increases to remove by-products of the metabolic exchange, which is captured using TCDS ( Aaslid, 1986 ). TCDS has become increasingly popular in cognitive neuroergonomic studies of vigilance and mental workload ( Warm and Parasuraman, 2007 ). However, because cerebral blood volume and blood flow velocity is influenced by systemic changes such as heart rate and blood pressure during exercise ( Ainslie et al., 2007 ), TCDS is less popular in assessing task-related neuronal activity in physical neuroergonomic studies of fatigue.

In contrast to the excellent temporal resolution offered by EEG techniques (on the order of milliseconds), magnetic resonance imaging (MRI) provides a structural image of the brain and offers excellent spatial visualization of deep internal parts, such as the hippocampus. While MRI provides static images of the brain that is critical in examining structural changes in the brain due to diseases (such as tumor), its application in studying structural changes in the brain over time (i.e., plasticity) has provided important information on learning and training ( Huttenlocher, 2002 ). A relatively newer MRI technique, called diffusion tensor imaging (DTI), uses MRI to target the diffusion of water molecules in the axons that make up white matter in the brain and allows for the computation of fractional anisotrophy (FA). FA values can range from 0 to 1, where 0 indicates non-directional (isotropic) and 1 indicates perfectly directional (anisotropic) diffusion. Higher FA values are thought to reflect greater integrity of white matter linking different cortical and subcortical regions of the brain. Several recent studies have assessed the effectiveness of cognitive and motor training on white matter integrity using the DTI technique ( Draganski et al., 2004 ; Takeuchi et al., 2010 ; Strenziok et al., 2014 ). In general, the MRI technique does not offer any mobility features, but an MRI static image can be overlaid with more dynamic fMRI images (i.e., blood oxygenation) so that areas of activation can be associated with particular brain regions.

The electromagnetic and hemodynamic neuroimaging techniques discussed thus far are based on sensing brain activity while a human operator is engaged in cognitive or physical work. As such, all such techniques are correlational , thus it may be difficult to establish causal links between brain activity and performance using these methods. Researchers have therefore turned to non-invasive stimulation techniques that modulate brain activity, such as transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS), in order to establish such causal associations. These techniques allow for temporary inhibition or activation of specific brain regions thereby allowing researchers to examine the causal role of different brain regions in various cognitive functions ( Walsh and Pascual-Leone, 2005 ). TMS and tDCS can also be used to modulate brain activity so that the performance of a given cognitive or motor task is improved ( Coffman et al., 2014 ). Alternatively, these techniques can also be applied not to enhance performance over baseline, but to reduce or eradicate a normally occurring performance limitation, such as performance decrements that occur in vigilance tasks ( Nelson et al., 2014 ).

Transcranial magnetic stimulation uses a magnetic coil that is positioned over the participant’s scalp over a brain region of interest to send electrical current that changes the magnetic field perpendicular to the head. This induces current flow in the underlying cortical issue, sufficient to alter neural firing ( Walsh and Pascual-Leone, 2005 ). The spatial resolution of TMS is relatively high, particularly when the participant’s MRI scan is available to guide the TMS coil placement. The temporal resolution is also high, since the TMS pulses can be delivered with millisecond precision. However, due to the equipment setup TMS does not offer a sufficient degree of mobility needed for neuroergonomic assessment in naturalistic work settings. Whereas TMS uses changing magnetic pulses, tDCS uses small DC electric current (1 or 2 mA) with electrodes attached to the scalp. A positive polarity (anode) is typically used to stimulate neuronal function and enhance performance, and a negative polarity (cathode) is used to inhibit neuronal activity. Compared to TMS, tDCS has low spatial and temporal resolution, but has the advantage that it is portable and very inexpensive and thus is more likely to be adopted in applied neuroergonomic studies.

Physical Work

Ergonomics began as the science of work to maximize productivity, particularly in physical work environments, but has since then expanded to become a scientific discipline concerned with the understanding of the interactions among humans and other elements of a system, in order to optimize human well-being and overall system performance. Physical ergonomics focuses on human physical capabilities and limitations, pertaining to anthropometry, physiology, and biomechanics of the human body, as they relate to physical work ( Karwowski et al., 2003 ). Traditional ergonomic evaluations focus solely on peripheral outcomes, such as force or muscle activity, and disregard the contributions of the brain during work. Physical neuroergonomics is an emerging field of study that focuses on the knowledge of human brain activities in relation to the control and design of physical tasks ( Karwowski et al., 2003 ), by taking into consideration an operator’s physical, cognitive, and affective capabilities and limitations. Here we consider how neuroergonomic methods have been employed to evaluate different physical work parameters (such as force production and repetition) and physical fatigue (localized muscle fatigue and whole body fatigue).

Physical Work Parameters

The primary goal of ergonomics is to ensure that work demands are always lower than operator capacity, and the conventional assessment of work demands include measuring biomechanical and physiological outcomes, such as joint torque, muscle activity, and heart rate, in laboratory and field settings. There has been recent interest in assessing physical work using neuroergonomic methods in controlled laboratory conditions; however, there is a clear lack of neuroergonomic studies in assessing physical work in actual field/work settings. Like any new field, physical neuroergonomics research first needs to understand the capabilities, limitations, and considerations of existing neuroimaging techniques on simulated work environments that can help build the knowledge base necessary to perform research in naturalistic work environments.

Since physical work can involve both static and dynamic work at different intensities, repetitions, and durations, which in turn can affect autonomic responses, different work parameters can influence the type of measurement technique adopted. For example, dynamic or ambulatory tasks, such as walking or lifting, cannot be assessed using fMRI due to mobility constraints. More appropriate neuroimaging methods to evaluate ambulatory physical work are EEG, ERP, and fNIRS. Of these, EEG appears to be the most common neuroimaging technique since it provides excellent temporal resolution. Effective artifact removal techniques are available that allow for its use in evaluating dynamic tasks. For example, EEG-derived MRCP has provided valuable information on the role of cortical motor commands (represented by the MRCP) on the control of voluntary muscle activation. MRCP from the supplementary motor area and the contralateral sensorimotor cortex has shown to be highly correlated with force production and rate of force production during isometric elbow-flexion, and associated muscle activity ( Siemionow et al., 2000 ). Of note, a recent fNIRS investigation has demonstrated obesity-related alterations in neural patterns of force control (i.e., lower prefrontal cortex activation associated with decreased joint stability) that can shed some light on the increased incidence of injury rates and higher work absenteeism in obese workers ( Schulte et al., 2007 ; Mehta and Shortz, 2013 ). High repetition is one of the major work-related risk factors that contribute to the development of musculoskeletal disorders ( Bernard, 1997 ). To evaluate the effects of repetition that involves flexion and extension of a joint, traditional ergonomic methods focus on muscular responses such as EMG. In a study investigating thumb flexion and extension movements, EEG-derived MRCP findings from the supplementary motor area and contralateral motor cortex demonstrated that extension and flexion result from separate corticospinal projections to the motor neurons ( Yue et al., 2000 ). Thumb extensions resulted in lower EMG but elicited greater brain responses than flexion movements. This particular finding may be important to our understanding of the etiology of musculoskeletal disorders due to repetitive motion. Real work environments are seldom static, and can require operators to focus not only on the physical work demands but also on the necessary visual/auditory cues associated with the tasks. Such tasks, which are dynamic and require visuomotor control, have shown to increase corticomuscular coupling at higher EEG frequencies (i.e., gamma bands), indicating the adaptive role of cortical oscillations in rapidly integrating visual (or new) information with the somatosensory information ( Marsden et al., 2000 ; Omlor et al., 2007 ). These findings have important implications for task analysis and design, particularly for work tasks that require visual feedback or fine or precise control of body motions.

Physical Fatigue

Fatigue is defined as the inability to maintain required power after prolonged use of the muscle(s) ( Latash et al., 2003 ), and can be affected by central (i.e., motivation, cortical activity, etc.) and peripheral (i.e., changes in muscle contractile properties) mechanisms. Neuroergonomic methods can help examine the role of central brain mechanisms in fatigue development. Based on the work tasks, fatigue in the workplace can be broadly categorized as localized muscle fatigue, which is the fatigue of specific muscle groups during tasks such as assembly line work or precision work, and whole body fatigue, which is more cardiorespiratory in nature that can occur during manual materials handling tasks. Commonly used ergonomic indicators of localized muscle fatigue include a reduction in force generating capacity ( Vøllestad, 1997 ) and a decrease in EMG power spectrum ( Mehta and Agnew, 2012 ). However these measures do not delineate the contributions of central fatigue from peripheral fatigue. Using EEG-derived MRCP, Johnston et al. (2001) demonstrated a significant increase in the activity of the BP component and the motor potential (MP) component of the MRCP, associated with a decline in force production and reduced EMG activity during a fatiguing grasping task. These increases in the early components of MRCP may reflect development of compensatory cortical strategies to accommodate for the inability to maintain the desired force levels due to peripheral fatigue. Supporting this, Liu et al. (2007) advocated that muscle fatigues well before the brain does; in essence that peripheral fatigue occurs before central fatigue. They demonstrated, by estimating the changes of source locations of high-density EEG signals using a single moving current dipole model, that handgrip muscle fatigue was associated with shifting of brain activation centers from one location to another when neurons in the previous location become fatigued. These studies collectively demonstrate the application of EEG in examining the neural correlates of localized fatigue development of smaller muscles during relatively static, or immobile, tasks.

Of the various neuroimaging techniques, EEG offers the greatest flexibility and mobility features that make it an attractive candidate in assessing whole body fatigue. By simultaneously obtaining information on eye movements and spontaneous EEG signals, Kubitz and Mott (1996) demonstrated increased brain activation (i.e., decreased alpha activity and increased beta activity) during a fatiguing cycling task. While technical advances have been made in minimizing mechanical artifacts from high-density EEG signals during whole body movements ( Gwin et al., 2010 ), fNIRS has gained rapid attention in evaluating whole body fatigue owing to its methodological advantages over EEG. First, fNIRS provides information on the location of the neural signal generated, whereas with EEG signals, source localization has to be computationally derived. Second, there are no time-sensitive requirements in examining whole body fatigue when compared to fast reaction time tasks; slower hemodynamic responses of fNIRS are thus appropriate when compared to fast EEG responses. As such, fNIRS responses have shown to be less affected by movement artifacts than EEG signals ( Perrey, 2008 ). Several fNIRS studies have reported a significant decrease in relative levels of oxygenated hemoglobin in the prefrontal cortex, accompanied by muscular impairment, at exhaustion during submaximal and maximal fatiguing contractions ( González-Alonso et al., 2004 ; Bhambhani et al., 2007 ; Nybo and Rasmussen, 2007 ). In particular, Thomas and Stephane (2008) demonstrated that oxygenated hemoglobin levels in the prefrontal cortex during incremental cycling exercise increased in the early stages, but decreased markedly in the last stage until exhaustion. These findings imply that prefrontal cortex activation is associated with reduction in motor output at the cessation of exercise. However, these fatiguing tasks are accompanied by cardiorespiratory changes in the autonomic system that can affect fNIRS responses ( Obrig et al., 1996 ). Depending on the research questions asked, such systemic influences on cerebral hemodynamic responses may be desired or undesired. Obrig and Villringer (2003) emphasize the importance of analyzing deoxygenated hemoglobin levels as an indicator of “neuronal activation” over the more commonly used oxygenated hemoglobin values. They argue that oxygenated hemoglobin levels are acceptable neuronal activity indicators when cerebral autoregulation is intact, i.e., cerebral blood flow is in homeostasis. Increases in oxygenated hemoglobin during exercise can be attributed not only to neuronal activation but also to exercise-induced increased blood flow to the brain, and as such a decrease in deoxygenated hemoglobin is the most valid parameter. Thus, neuroergonomic investigations of fatigue need to consider these systemic influences, and perhaps collect peripheral measurements such as arterial blood pressure and heart rate to ensure that appropriate inferences are made from fNIRS signals.

Cognitive Work

The field of human factors and ergonomics had its origins in time-and-motion studies conducted in the early 1900s. With the advent of World War II, increasing attention was paid to evaluation of human psychological processes during work performance, but the dominant approach was behaviorism, or stimulus-response psychology. The advent of the cognitive revolution in the late 1950s lead to the introduction of the cognitive approach in human performance assessment from the 1960s to the present day, but there was still a relative neglect of brain mechanisms. Advances in neuroimaging and related methods that lead to the development of the field of cognitive neuroscience lead to the argument that neural measures should also be considered in human factors and ergonomics ( Parasuraman, 2003 ). Since that time, the neuroergonomic approach has been applied to a number of different issues in cognitive ergonomics.

These historical trends in theoretical frameworks used in ergonomics can be seen clearly in the periodical reviews of the field of engineering psychology in the Annual Reviews of Psychology . Fitts (1958) reviewed work conducted mainly within time-and-motion and stimulus-response frameworks; Wickens and Kramer (1985) presented a cognitive or information-processing approach; and the most recent review, by Proctor and Vu (2010) , describes the neuroergonomic approach. In this paper, we review a few key issues in cognitive neuroergonomics and on those areas where the most research and development work has been done. These include: (1) mental workload, (2) vigilance and mental fatigue, and (3) neuroadaptive systems.

Mental Workload

The assessment of human mental workload is one of the most widely studied topics in ergonomics ( Wickens and McCarley, 2008 ). If operator mental workload is either too high or too low human-system performance may suffer in work environments, thereby potentially compromising safety. Hence, workload must be assessed in the design of new systems or the evaluation of existing ones. Behavioral measures, such as accuracy and speed of response on secondary tasks, or subjective reports (such as the NASA-TLX) have been widely used to assess mental workload. However, measures of brain function offer some unique advantages that can be exploited in mental workload assessment ( Kramer and Parasuraman, 2007 ). Among these is the ability to extract covert physiological measures continuously in complex system operations in which overt behavioral measures may be relatively sparse.

The dominant theory of human mental workload is resource theory ( Wickens, 1984 , 2002 ). This theory postulates that except for highly overlearned “automatic” tasks, task performance is directly proportional to the application of attentional resources. The theory also proposes that the degree of overlap of multiple pools of resources determines the pattern and amount of interference when two or more tasks are performed simultaneously (such as driving and talking on the cell phone). Dual-task studies have provided abundant support for resource theory ( Wickens and McCarley, 2008 ), but one criticism is that the theory is circular ( Navon, 1984 ), which can be linked to the lack of an independent measure of resources. This criticism can be countered if neural measures of mental resources can be identified.

Measures of cerebral hemodynamics, such as fNIRS and TCDS, have provided validation for the resource construct. In a recent study, Ayaz et al. (2012) tested experienced air traffic controllers (ATC) on a complex ATC task requiring them to keep aircraft in their sector free of conflicts. fNIRS was used to measure activation of the prefrontal cortex. Ayaz et al. (2012) found there was an increase in prefrontal cortex activation as the number of aircraft in their sector increased. These findings suggest that fNIRS can provide a sensitive index of cognitive workload in a skilled group performing a realistic task that was highly representative of their work environment. fNIRS has also been found to index changes in prefrontal cortex activation with skill acquisition in both basic working memory tasks ( McKendrick et al., 2014 ) and more complex piloting tasks ( Ayaz et al., 2012 ). Most recently, portable versions of fNIRS have been developed for use in mobile neuroimaging ( Ayaz et al., 2013 ).

There are many factors, such as cost, ease of implementation, intrusiveness, etc., that must be taken into consideration when selecting neuroergonomic techniques for mental workload assessment. Some of these factors (e.g., cost) may rule out the use of neuroergonomic methods in favor of simpler indexes such as subjective measures. Some workers may also not wish to be “wired up” for physiological recording, so operator acceptance must also be carefully considered. However, with increasing miniaturization and development of dry electrode, wireless, wearable systems, some of these concerns are diminishing.

Vigilance and Mental Fatigue

The evaluation of operator vigilance and mental fatigue in work environments is a topic closely related to workload assessment. The widespread implementation of automation in many work environments, including air and surface transportation and health care, while often leading to a reduction in operator workload, can also increase workload because of the resulting need for monitoring the automation ( Parasuraman, 1987 ). The typical finding in vigilance studies is that the detection rate of critical targets declines with time on task ( Davies and Parasuraman, 1982 ). Vigilance decrement was originally attributed to a reduction in physiological arousal ( Frankmann and Adams, 1962 ) but more recent neuroergonomic research using TCDS and fNIRS have attributed it to resource depletion ( Warm et al., 2008 ). Warm et al. (2008) reported a series of studies of TCDS and vigilance (for reviews, see Warm and Parasuraman, 2007 ; Warm et al., 2008 ). A consistent finding is that the vigilance decrement is paralleled by a decline in blood flow velocity over time, relative to a baseline of activity just prior to beginning the vigilance session. The parallel decline in vigilance performance and in blood flow velocity is found for both visual and auditory tasks ( Shaw et al., 2009 ). These findings have been interpreted using resource theory. A critical control finding in support of resource theory – as opposed to a generalized arousal or fatigue model – is that the blood flow change occurs only when observers actively engage with the vigilance task. When observers are asked to simply watch a display passively without having to detect targets for the same amount of time as in an active vigilance condition – a case of maximal under-arousal – blood flow velocity does not decline but remains stable over time.

The deleterious effects of loss of operator vigilance can countered with reduced work hours and more frequent rest breaks, but this may not be practical in all work settings. Another mitigating strategy is to use cueing. Detection performance in vigilance tasks can be improved by providing observers with consistent and reliable cues to the imminent arrival of critical signals, with the extent of the decrement being reduced or eliminated ( Wiener and Attwood, 1968 ). With cueing, observers need to monitor a display only after having been prompted about the arrival of a signal and therefore can husband their information processing resources over time. In contrast, when no cues are provided, observers are never certain of when a critical signal might appear and consequently have to process information on their displays continuously across the watch, thereby consuming more of their resources over time than cued observers. If the vigilance decrement stems from resource depletion due to need to attend continuously to a display, then pre-cues should reduce the decline in cerebral blood flow velocity as measured by TCDS. This was confirmed in a study by Hitchcock et al. (2003) . They used no pre-cues or pre-cues that were 100, 80, or 40% reliable in pointing to an upcoming critical event in a simulated air traffic control task. Performance efficiency remained stable when perfectly reliable cues were provides but declined over time in the remaining conditions, so that by the end of the vigil, performance efficiency was clearly best in the 100% group, followed in order by the 80, 40%, and no-cue groups. Blood flow declined in the no cue control condition, but there was a progressive reduction in the extent of the decline with progressively more reliable cues. There was no decline when the cues were perfectly reliable. This pattern of change in blood flow exactly matched that of performance.

In addition to cueing, non-invasive brain stimulation could also be used to mitigate vigilance decrement and mental fatigue. Nelson et al. (2014) applied 1 mA anodal tDCS to either the left or right prefrontal cortex while participants performed the same vigilance task used by Hitchcock et al. (2003) . tDCS was applied either early or late during the course of the vigilance task. Compared to a control group that showed the normal vigilance decrement, the early stimulation group had a higher detection rate of critical signals. The late stimulation group initially exhibited a vigilance decrement, but this was reversed following application of tDCS. These initial findings are highly encouraging, but need to be followed up with additional research to examine the long-term effectiveness of tDCS as a method to alleviate vigilance problems at work.

Training and Neuroadaptive Systems

While the goal of ergonomic design is to avoid having workers exposed to extremes of workload and to loss of vigilance, this may not always be possible in certain work settings where unexpected events, equipment failures, or other unanticipated factors lead to a transient increase in the task load imposed on the human operator, or long work hours impose demands on operator vigilance. Adaptive automation offers one approach to deal with these issues ( Parasuraman, 1987 , 2000 ). In this approach, the allocation of functions to human and machine agents is flexible during system operations, with greater use of automation during high task load conditions or emergencies and less during normal operations, consistent with the approach of dynamic function allocation ( Lintern, 2012 ). The adaptive automation concept has a long history ( Parasuraman et al., 1992 ), but neuroergonomic methods for its implementation have been considered relatively recently ( Inagaki, 2003 ; Parasuraman, 2003 ; Scerbo, 2007 ).

Several methods to implement adaptive systems have been examined, including neuroergonomic measures to assess the operator’s functional state ( Byrne and Parasuraman, 1996 ; Kramer and Parasuraman, 2007 ; Wilson and Russell, 2007 ; Parasuraman and Wilson, 2008 ; Ting et al., 2010 ). Many studies have used EEG because of its ease of recording and (relative) unobtrusiveness (compared, say, to secondary tasks or subjective questionnaires). EEG also has the property of being a very high bandwidth measure, offering the possibility of sampling the human operator at up to about 30 Hz ( Wilson and Russell, 2003 ). Workload adaptive systems need to assess operator state in real time, or near real time, so that task allocation or restructuring can be implemented in cases of overload or underload. A number of different statistical and machine learning techniques have been used for this purpose. These include discriminant analysis ( Berka et al., 2004 ), artificial neural networks ( Wilson and Russell, 2007 ; Baldwin and Penaranda, 2012 ), Bayesian networks ( Wang et al., 2011 ), and fuzzy logic ( Ting et al., 2010 ). These have been implemented in real time and typically provide accuracies of 70–85%.

Implementing neuroergonomic adaptive systems in real settings poses significant challenges. A major issue concerns the detection and removal of artifacts in real time. Furthermore, while initial success has been achieved in using computational techniques to classify workload on the basis of EEG and other neuroergonomic measures, the reliability and stability of these methods within and across individuals needs to be more rigorously tested ( Wang et al., 2011 ; Christensen et al., 2012 ). Finally, the operational community must be involved in the design of adaptive systems to ensure user acceptance and compliance.

Neuroergonomic Assessment of Concurrent Physical and Cognitive Work

Both physical and cognitive neuroergonomics have helped advance our understanding on the role of the human brain during physical and cognitive work, respectively. Only a small number of studies have investigated the interaction between physical and cognitive work, which is a big concern since “work” places combined physical and cognitive demands on operators, never either one in isolation. High cognitive demands can influence physical work; and physical activity can in turn influence cognitive processing. In comparison to traditional evaluation techniques in either physical or cognitive ergonomics domain, neuroergonomic methods offer a great advantage in assessing these combined demands. For example, using EEG signals Kamijo et al. (2004) investigated the influence of exercise intensity on cognitive function using the P300 ERP component. They suggested that exercise influenced the amount of attentional resources devoted to a given task and that the changes in P300 amplitude followed an inverted U -shaped behavior of differences in exercise intensity. When examining the impact of cognitive demand on physical capacity, a few studies have attributed decreased muscle endurance in presence of a cognitively stressful situation to lower motivation ( Marcora et al., 2009 ), increased neuromotor noise impairing joint steadiness ( van Loon et al., 2001 ; Mehta and Agnew, 2011 , 2013 ; Mehta et al., 2012 ), or neuronal interference at the prefrontal cortex that is involved in cognitive processing and isometric motor contractions ( Dettmers et al., 1996 ; Rowe et al., 2000 ; Mehta and Parasuraman, 2013 ). In particular, using fNIRS to monitor cerebral oxygenation during handgrip exercises, Mehta and Parasuraman (2013) demonstrated that concurrent handgrip exercises in cognitive stressful conditions were associated with lower oxygenated hemoglobin levels in the bilateral prefrontal cortex at exhaustion when compared to the handgrip exercises at the same intensity levels (i.e., no changes observed in peripheral muscular responses of EMG and force exerted) under no stress. Quite similarly, using EEG- and EMG-derived corticomuscular coupling measure, Kristeva-Feige et al. (2002) reported that corticomuscular coupling decreased significantly during a cognitively stressful condition despite no changes observed in traditional measures such as EMG and force production. These studies collectively emphasize the importance of obtaining brain (or central) responses along with the more conventional ergonomic measurements to accurately understand the “total” demands placed on humans during work that requires both physical and cognitive processing. Future investigations on comparing these brain-body responses with the more traditional performance or subjective measures are also needed to understand the underlying neural “cost” of operator functional state. Such studies are also needed so as to develop evaluation tools (surveys, heuristic checklists) that are predictive of the neural and physiological cost associated with optimizing work tasks, which can be used by designers or supervisors to quantify operator workload and fatigue.

Mobile Brain Imaging Considerations for Workload/Fatigue Assessments

One of the key distinctions between neuroergonomics and neuroscience is that neuroergonomics is the study of brain and behavior “at work.” Thus, it is extremely important that neuroergonomic methods are capable of examining human operators at their naturalistic work settings. In this paper, we discussed the merits and disadvantages of the available neuroimaging techniques applicable to neuroergonomics and a key theme identified was the lack of studies evaluating neural bases of mobile work, particularly in the physical neuroergonomics domain. Recent efforts in developing mobile brain imaging (MoBI) techniques, which consider the physical and environmental impact on human cognitive processing, show great promise. For example, Gramann et al. (2011) reviewed the implications and feasibility of a newly developed MoBI system that was previously employed in examining cognitive processing during human stance and locomotion. In particular, their MoBI investigation included simultaneous brain-body measurements from a 256-channel EEG system and kinematic and kinetic outcomes that are otherwise employed during conventional gait biomechanics using motion capture systems and force plates ( Gwin et al., 2011 ). In their review, Gramann and colleagues identify key requirements for MoBI methods that include: (1) robust mobile sensor technology to measure brain activity, (2) comprehensive “wireless” body measurement system, and (3) powerful computational software to collectively processing and analyze both brain-body responses. While developing an ideal MoBI system may be a challenging goal, understanding current limitations in mobile brain-body imaging and addressing them, albeit painstakingly, is a critical step toward achieving this goal. Future investigations can also include developing similar mobile brain-body imaging systems for hemodynamic neuroimaging techniques, utilizing either fNIRS or TCDS to provide brain imaging measures, and using peripheral measurements such as heart rate and blood pressure to document physiological whole-body responses.

Ergonomics has long since moved from being a science of improving work efficiency to now being focused on enhancing well-being while improving systems performance. To effectively understand how humans interact with work systems, it is not only important to ask how well they perform, but also why they perform a certain way. Neuroergonomics have helped fill in the gaps on the neural bases of both physical and cognitive performance that were left unanswered with traditional ergonomic assessments. In this review we discussed the recent developments and adoption of neuroergonomic methods and applications in investigating physical, cognitive, and combined physical and cognitive work. We also reviewed the applicability and feasibility of neuroimaging techniques in evaluating mobile work environments. While some neuroimaging methods are expensive and are immobile, such as the MRI, fMRI, PET, and DTI, portable methods such as EEG, fNIRS, and TCDS, are more likely to be adopted in applied ergonomics research. With the advent of, and recent developments in, MoBI technology, we can be assured that neuroergonomics can continue providing critical information on how/why human interact in ambulatory and naturalistic work settings.

Author Contributions

Both authors contributed equally to this work. Ranjana K. Mehta performed the literature review on neuroergonomics applications to physical work and Raja Parasuraman performed the literature review on cognitive neuroergonomics. Both authors discussed the reviewed implications and commented on the manuscript at all stages.

Conflict of Interest Statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Acknowledgments

Supported in part by Air Force Office of Scientific Research grant FA9550-10-1-0385 to Raja Parasuraman.

Aaslid, R. (1986). “Transcranial Doppler examination techniques,” in Transcranial Doppler Sonography , ed. R. Aaslid (New York: Springer-Verlag), 39–59.

Ainslie, P. N., Barach, A., Murrell, C., Hamlin, M., Hellemans, J., and Ogoh, S. (2007). Alterations in cerebral autoregulation and cerebral blood flow velocity during acute hypoxia: rest and exercise. Am. J Physiol. Heart Circ. Physiol. 292, H976–H983. doi: 10.1152/ajpheart.00639.2006

Pubmed Abstract | Pubmed Full Text | CrossRef Full Text

Ayaz, H., Onaral, B., Izzetoglu, K., Shewokis, P. A., McKendrick, R., and Parasuraman, R. (2013). Continuous monitoring of brain dynamics with functional near infrared spectroscopy as a tool for neuroergonomic research: empirical examples and a technological development. Front. Hum. Neurosci. 7:871. doi: 10.3389/fnhum.2013.00871

CrossRef Full Text

Ayaz, H., Shewokis, P. A., Bunce, S., Izzetoglu, K., Willems, B., and Onaral, B. (2012). Optical brain monitoring for operator training and mental workload assessment. Neuroimage 59, 36–47. doi: 10.1016/j.neuroimage.2011.06.023

Baldwin, C. L., and Penaranda, B. (2012). Adaptive training using an artificial neural network and EEG metrics for within-and cross-task workload classification. Neuroimage 59, 48–56. doi: 10.1016/j.neuroimage.2011.07.047

Berka, C., Levendowski, D. J., Cvetinovic, M. M., Petrovic, M. M., Davis, G., Lumicao, M. N., et al. (2004). Real-time analysis of EEG indexes of alertness, cognition, and memory acquired with a wireless EEG headset. Int. J. Hum. Comput. Interact. 17, 151–170. doi: 10.1207/s15327590ijhc1702_3

Bernard, B. P. (1997). Musculoskeletal Disorders and Workplace Factors: a Critical Review of Epidemiologic Evidence for Work-Related Musculoskeletal Disorders of the Neck, Upper Extremity, and Low Back . Cincinnati, OH: U.S. Department of Health and Human Services CDC (NIOSH), Publication No. 97–141.

Bhambhani, Y., Malik, R., and Mookerjee, S. (2007). Cerebral oxygenation declines at exercise intensities above the respiratory compensation threshold. Respir. Physiol. Neurobiol. 156, 196–202. doi: 10.1016/j.resp.2006.08.009

Byrne, E. A., and Parasuraman, R. (1996). Psychophysiology and adaptive automation. Biol. Psychol. 42, 249–268. doi: 10.1016/0301-0511(95)05161-9

Christensen, J. C., Estepp, J. R., Wilson, G. F., and Russell, C. A. (2012). The effects of day-to-day variability of physiological data on operator functional state classification. Neuroimage 59, 57–63. doi: 10.1016/j.neuroimage.2011.07.091

Clark, A. (1998). Being There: Putting Brain, Body, and World Together Again . Cambridge, MA: MIT Press.

Coffman, B. A., Clark, V. P., and Parasuraman, R. (2014). Battery powered thought: enhancement of attention, learning, and memory in healthy adults using transcranial direct current stimulation. Neuroimage 85, 895–908. doi: 10.1016/j.neuroimage.2013.07.083

Cui, X., Bray, S., Bryant, D. M., Glover, G. H., and Reiss, A. L. (2011). A quantitative comparison of NIRS and fMRI across multiple cognitive tasks. Neuroimage 54, 2808–2821. doi: 10.1016/j.neuroimage.2010.10.069

Davies, D. R., and Parasuraman, R. (1982). The Psychology of Vigilance . London: Academic Press.

Dettmers, C., Lemon, R., Stephen, K., Fink, G., and Frackowiak, R. (1996). Cerebral activation during the exertion of sustained static force in man. Neuroreport 7, 2103. doi: 10.1097/00001756-199609020-00008

Draganski, B., Gaser, C., Busch, V., Schuierer, G., Bogdahn, U., and May, A. (2004). Neuroplasticity: changes in grey matter induced by training. Nature 427, 311–312. doi: 10.1038/427311a

Duschek, S., and Schandry, R. (2003). Functional transcranial Doppler sonography as a tool in psychophysiological research. Psychophysiology 40, 436–454. doi: 10.1111/1469-8986.00046

Fazli, S., Mehnert, J., Steinbrink, J., Curio, G., Villringer, A., Müller, K. R., et al. (2012). Enhanced performance by a hybrid NIRS–EEG brain computer interface. Neuroimage 59, 519–529. doi: 10.1016/j.neuroimage.2011.07.084

Fedota, J. R., and Parasuraman, R. (2010). Neuroergonomics and human error. Theor. Issues Ergon. Sci. 11, 402–421. doi: 10.1080/14639220902853104

Fitts, P. M. (1958). Engineering psychology. Annu. Rev. Psychol. 9, 267–294. doi: 10.1146/annurev.ps.09.020158.001411

Frankmann, J. P., and Adams, J. (1962). Theories of vigilance. Psychol. Bull. 59, 257. doi: 10.1037/h0046142

González-Alonso, J., Dalsgaard, M. K., Osada, T., Volianitis, S., Dawson, E. A., Yoshiga, C. C., et al. (2004). Brain and central haemodynamics and oxygenation during maximal exercise in humans. J. Physiol. 557, 331–342. doi: 10.1113/jphysiol.2004.060574

Gramann, K., Gwin, J. T., Ferris, D. P., Oie, K., Jung, T. P., Lin, C. T., et al. (2011). Cognition in action: imaging brain/body dynamics in mobile humans. Rev. Neurosci . 22, 593–608. doi: 10.1515/RNS.2011.047

Gwin, J. T., Gramann, K., Makeig, S., and Ferris, D. P. (2010). Removal of movement artifact from high-density EEG recorded during walking and running. J. Neurophysiol. 103, 3526–3534. doi: 10.1152/jn.00105.2010

Gwin, J. T., Gramann, K., Makeig, S., and Ferris, D. P. (2011). Electrocortical activity is coupled to gait cycle phase during treadmill walking. Neuroimage 54, 1289–1296. doi: 10.1016/j.neuroimage.2010.08.066

Halliday, D., Rosenberg, J., Amjad, A., Breeze, P., Conway, B., and Farmer, S. (1995). A framework for the analysis of mixed time series/point process data-theory and application to the study of physiological tremor, single motor unit discharges and electromyograms. Prog. Biophys. Mol. Biol. 64, 237–278. doi: 10.1016/S0079-6107(96)00009-0

Hitchcock, E. M., Warm, J. S., Matthews, G., Dember, W. N., Shear, P. K., Tripp, L. D., et al. (2003). Automation cueing modulates cerebral blood flow and vigilance in a simulated air traffic control task. Theor. Issues Ergon. Sci. 4, 89–112. doi: 10.1080/14639220210159726

Huttenlocher, P. R. (2002). Neural Plasticity: The Effects of Environment on the Development of the Cerebral Cortex . Cambridge, MA: Harvard University Press.

Inagaki, T. (2003). “Adaptive automation: sharing and trading of control,” in Handbook of Cognitive Task Design , ed. E. Hollnagel (Mahwah, NJ: Lawrence Erlbaum), 46–89.

Jobsis, F. F. (1977). Noninvasive, infrared monitoring of cerebral and myocardial oxygen sufficiency and circulatory parameters. Science 198, 1264–1267. doi: 10.1126/science.929199

Johnston, J., Rearick, M., and Slobounov, S. (2001). Movement-related cortical potentials associated with progressive muscle fatigue in a grasping task. Clin. Neurophysiol. 112, 68–77. doi: 10.1016/S1388-2457(00)00452-1

Jung, T. P., Makeig, S., Humphries, C., Lee, T. W., Mckeown, M. J., Iragui, V., et al. (2000). Removing electroencephalographic artifacts by blind source separation. Psychophysiology 37, 163–178. doi: 10.1111/1469-8986.3720163

Kamijo, K., Nishihira, Y., Hatta, A., Kaneda, T., Wasaka, T., Kida, T., et al. (2004). Differential influences of exercise intensity on information processing in the central nervous system. Eur. J. Appl. Physiol. 92, 305–311. doi: 10.1007/s00421-004-1097-2

Karwowski, W., Siemionow, W., and Gielo-Perczak, K. (2003). Physical neuroergonomics: the human brain in control of physical work activities. Theor. Issues Ergon. Sci. 4, 175–199. doi: 10.1080/1463922021000032339

Kornhuber, H. H., and Deecke, L. (1965). Changes in the brain potential in voluntary movements and passive movements in man: readiness potential and reafferent potentials. Pflügers Arch. Gesamte Physiol. Menschen Tiere 284, 1–17. doi: 10.1007/BF00412364

Kramer, A., and Parasuraman, R. (2007). “Neuroergonomics – application of neuroscience to human factors,” in Handbook of Psychophysiology , 2nd Edn, eds J. Caccioppo, L. Tassinary, and G. Berntson (New York: Cambridge University Press), 704–722.

Pubmed Abstract | Pubmed Full Text

Kristeva-Feige, R., Fritsch, C., Timmer, J., and Lücking, C. H. (2002). Effects of attention and precision of exerted force on beta range EEG-EMG synchronization during a maintained motor contraction task. Clin. Neurophysiol. 113, 124–131. doi: 10.1016/S1388-2457(01)00722-2

Kubitz, K., and Mott, A. (1996). EEG power spectral densities during and after cycle ergometer exercise. Res. Q. Exerc. Sport 67, 91. doi: 10.1080/02701367.1996.10607929

Latash, M. L., Danion, F., and Bonnard, M. (2003). Effects of transcranial magnetic stimulation on muscle activation patterns and joint kinematics within a two-joint motor synergy. Brain Res. 961, 229–242. doi: 10.1016/S0006-8993(02)03958-6

Leamy, D. J., Collins, R., and Ward, T. E. (2011). “Combining fNIRS and EEG to improve motor cortex activity classification during an imagined movement-based task,” in Foundations of Augmented Cognition. Directing the Future of Adaptive Systems, eds D. D. Schmorrow and C. M. Fidopiastis (Berlin: Springer), 177–185.

Lintern, G. (2012). Work-focused analysis and design. Cogn. Technol. Work 14, 71–81. doi: 10.1007/s10111-010-0167-y

Liu, Z., and He, B. (2008). fMRI–EEG integrated cortical source imaging by use of time-variant spatial constraints. Neuroimage 39, 1198–1214. doi: 10.1016/j.neuroimage.2007.10.003

Liu, J. Z., Lewandowski, B., Karakasis, C., Yao, B., Siemionow, V., Sahgal, V., et al. (2007). Shifting of activation center in the brain during muscle fatigue: an explanation of minimal central fatigue? Neuroimage 35, 299–307. doi: 10.1016/j.neuroimage.2006.09.050

Luck, S. J. (2005). An Introduction to the Event-Related Potential Technique . Cambridge, MA: MIT Press.

Makeig, S., Gramann, K., Jung, T.-P., Sejnowski, T., and Poizner, H. (2009). Linking brain, mind and behavior. Int. J. Psychophysiol. 73, 95–100. doi: 10.1016/j.ijpsycho.2008.11.008

Marcora, S. M., Staiano, W., and Manning, V. (2009). Mental fatigue impairs physical performance in humans. J. Appl. Physiol. 106, 857–864. doi: 10.1152/japplphysiol.91324.2008

Marsden, J., Werhahn, K., Ashby, P., Rothwell, J., Noachtar, S., and Brown, P. (2000). Organization of cortical activities related to movement in humans. J. Neurosci. 20, 2307–2314.

McKendrick, R., Ayaz, H., Olmstead, R., and Parasuraman, R. (2014). Enhancing dual-task performance with verbal and spatial working memory training: continuous monitoring of cerebral hemodynamics with NIRS. Neuroimage 85, 1014–1026. doi: 10.1016/j.neuroimage.2013.05.103

Mehta, R. K., and Agnew, M. J. (2011). Effects of concurrent physical and mental demands for a short duration static task. Int. J. Ind. Ergon. 41, 488–493. doi: 10.1016/j.ergon.2011.04.005

Mehta, R. K., and Agnew, M. J. (2012). Influence of mental workload on muscle endurance, fatigue, and recovery during intermittent static work. Eur. J. Appl. Physiol. 112, 2891–2902. doi: 10.1007/s00421-011-2264-x

Mehta, R. K., and Agnew, M. J. (2013). Exertion-dependent effects of physical and mental workload on physiological outcomes and task performance. IIE Trans. Occupat. Ergon. Hum. Factors 1, 3–15. doi: 10.1080/21577323.2011.632488

Mehta, R. K., Nussbaum, M. A., and Agnew, M. J. (2012). Muscle- and task-dependent responses to concurrent physical and mental workload during intermittent static work. Ergonomics 59, 1166–1179. doi: 10.1080/00140139.2012.703695

Mehta, R. K., and Parasuraman, R. (2013). Effects of mental fatigue on the development of physical fatigue: a neuroergonomics approach. Hum. Factors doi: 10.1177/0018720813507279 [Epub ahead of print].

Mehta, R. K., and Shortz, A. E. (2013). Obesity-related differences in neural correlates of force control. Eur. J. Appl. Physiol. doi: 10.1007/s00421-013-2762-0 [Epub ahead of print].

Navon, D. (1984). Resources – a theoretical soup stone? Psychol. Rev. 91, 216. doi: 10.1037/0033-295X.91.2.216

Nelson, J. T., Mckinley, R. A., Golob, E. J., Warm, J. S., and Parasuraman, R. (2014). Enhancing vigilance in operators with prefrontal cortex transcranial direct current stimulation. Neuroimage 85, 907–917. doi: 10.1016/j.neuroimage.2012.11.061

Nybo, L., and Rasmussen, P. (2007). Inadequate cerebral oxygen delivery and central fatigue during strenuous exercise. Exerc. Sport Sci. Rev. 35, 110. doi: 10.1097/jes.0b013e3180a031ec

Obrig, H., Hirth, C., Junge-Hulsing, J., Doge, C., Wolf, T., Dirnagl, U., et al. (1996). Cerebral oxygenation changes in response to motor stimulation. J. Appl. Physiol. 81, 1174.

Obrig, H., and Villringer, A. (2003). Beyond the visible – imaging the human brain with light. J. Cereb. Blood Flow Metab. 23, 1–18. doi: 10.1097/00004647-200301000-00001

Omlor, W., Patino, L., Hepp-Reymond, M. C., and Kristeva, R. (2007). Gamma-range corticomuscular coherence during dynamic force output. Neuroimage 34, 1191–1198. doi: 10.1016/j.neuroimage.2006.10.018

Parasuraman, R. (1987). Human-computer monitoring. Hum. Factors 29, 695–706.

Parasuraman, R. (2000). Designing automation for human use: empirical studies and quantitative models. Ergonomics 43, 931–951. doi: 10.1080/001401300409125

Parasuraman, R. (2003). Neuroergonomics: research and practice. Theor. Issues Ergon. Sci. 4, 5–20. doi: 10.1080/14639220210199753

Parasuraman, R. (2011). Neuroergonomics: brain, cognition, and performance at work. Curr. Dir. Psychol. Sci. 20, 181–186. doi: 10.1177/0963721411409176

Parasuraman, R., Bahri, T., Deaton, J. E., Morrison, J. G., and Barnes, M. (1992). Theory and Design of Adaptive Automation in Aviation Systems . Warminster, PA: Naval Air Development Center.

Parasuraman, R., and Rizzo, M. (2007). Neuroergonomics: The Brain at Work . Oxford: Oxford University Press.

Parasuraman, R., and Wilson, G. F. (2008). Putting the brain to work: neuroergonomics past, present, and future. Hum. Factors 50, 468–474. doi: 10.1518/001872008X288349

Perrey, S. (2008). Non-invasive NIR spectroscopy of human brain function during exercise. Methods 45, 289–299. doi: 10.1016/j.ymeth.2008.04.005

Pizzagalli, D. A. (2007). “Electroencephalography and high-density electrophysiological source localization,” in Handbook of Psychophysiology , 3rd Edn, eds. J. T. Cacioppo, L. G. Tassinary, and G. G. Bernston (Cambridge: Cambridge University Press), 56–84.

Poldrack, R., Mumford, J. A., and Nichols, T. E. (2011). Handbook of Functional MRI Analysis . New York: Cambridge University Press. doi: 10.1017/CBO9780511895029

Proctor, R. W., and Vu, K. L. (2010). Cumulative knowledge and progress in human factors. Annu. Rev. Psychol. 61, 623–651. doi: 10.1146/annurev.psych.093008.100325

Raz, A., Lieber, B., Soliman, F., Buhle, J., Posner, J., Peterson, B. S., et al. (2005). Ecological nuances in functional magnetic resonance imaging (fMRI): psychological stressors, posture, and hydrostatics. Neuroimage 25, 1–7. doi: 10.1016/j.neuroimage.2004.11.015

Rowe, J., Toni, I., Josephs, O., Frackowiak, R., and Passingham, R. (2000). The prefrontal cortex: response selection or maintenance within working memory? Science 288, 1656. doi: 10.1126/science.288.5471.1656

Scerbo, M. (2007). “Adaptive automation,” in Neuroergonomics: The Brain at Work , eds R. Parasuraman and M. Rizzo (New York: Oxford University Press), 239–252.

Schulte, P., Wagner, G., Ostry, A., Blanciforti, L., Cutlip, G., and et al. (2007). Work, obesity, and occupational safety and health. Am. J. Public Health 97, 428–436. doi: 10.2105/AJPH.2006.086900

Shaw, T. H., Warm, J. S., Finomore, V., Tripp, L., Matthews, G., Weiler, E., et al. (2009). Effects of sensory modality on cerebral blood flow velocity during vigilance. Neurosci. Lett. 461, 207–211. doi: 10.1016/j.neulet.2009.06.008

Siemionow, V., Yue, G. H., Ranganathan, V. K., Liu, J. Z., and Sahgal, V. (2000). Relationship between motor activity-related cortical potential and voluntary muscle activation. Exp. Brain Res. 133, 303–311. doi: 10.1007/s002210000382

Strangman, G., Culver, J. P., Thompson, J. H., and Boas, D. A. (2002). A quantitative comparison of simultaneous BOLD fMRI and NIRS recordings during functional brain activation. Neuroimage 17, 719–731. doi: 10.1006/nimg.2002.1227

Strenziok, M., Parasuraman, R., Clarke, E., Cisler, D., Thompson, J. C., and Greenwood, P. M. (2014). Neurocognitive enhancement in older adults: comparison of three cognitive training tasks to test a hypothesis of training transfer in brain connectivity. Neuroimage 85, 1027–1039. doi: 10.1016/j.neuroimage.2013.07.069

Takeuchi, H., Sekiguchi, A., Taki, Y., Yokoyama, S., Yomogida, Y., Komuro, N., et al. (2010). Training of working memory impacts structural connectivity. J. Neurosci. 30, 3297–3303. doi: 10.1523/JNEUROSCI.4611-09.2010

Thomas, R., and Stephane, P. (2008). Prefrontal cortex oxygenation and neuromuscular responses to exhaustive exercise. Eur. J. Appl. Physiol. 102, 153–163. doi: 10.1007/s00421-007-0568-7

Ting, C. H., Mahfouf, M., Nassef, A., Linkens, D. A., Panoutsos, G., Nickel, P., et al. (2010). Real-time adaptive automation system based on identification of operator functional state in simulated process control operations. IEEE Trans. Syst. Man Cybern. Syst. Hum. 40, 251–262. doi: 10.1109/TSMCA.2009.2035301

van Loon, E. M., Masters, R. S. W., Ring, C., and Mcintyre, D. B. (2001). Changes in limb stiffness under conditions of mental stress. J. Mot. Behav. 33, 153–164. doi: 10.1080/00222890109603147

Villringer, A., and Chance, B. (1997). Non-invasive optical spectroscopy and imaging of human brain function. Trends Neurosci. 20, 435–442. doi: 10.1016/S0166-2236(97)01132-6

Vøllestad, N. K. (1997). Measurement of human muscle fatigue. J. Neurosci. Methods 74, 219–227. doi: 10.1016/S0165-0270(97)02251-6

Walsh, V., and Pascual-Leone, A. (2005). Transcranial Magnetic Stimulation: A Neurochronometrics of Mind . Cambridge, MA: Bradford.

Wang, Z., Hope, R. M., Wang, Z., Ji, Q., and Gray, W. D. (2011). Cross-subject workload classification with a hierarchical Bayes model. Neuroimage 59, 64–69. doi: 10.1016/j.neuroimage.2011.07.094

Warm, J. S., and Parasuraman, R. (2007). “Cerebral hemodynamics and vigilance,” in Neuroergonomics: The Brain at Work , eds R. Parasuraman and M. Rizzo (New York: Oxford University Press), 146–158.

Warm, J. S., Parasuraman, R., and Matthews, G. (2008). Vigilance requires hard mental work and is stressful. Hum. Factors 50, 433–441. doi: 10.1518/001872008X312152

Wickens, C. D. (1984). “Processing resources in attention,” in Varieties of Attention , eds R. Parasuraman and R. Davies (Orlando, FL: Academic Press), 63–101.

Wickens, C. D. (2002). Multiple resources and performance prediction. Theor. Issues Ergon. Sci. 3, 159–177. doi: 10.1080/14639220210123806

Wickens, C. D., and Kramer, A. (1985). Engineering psychology. Annu. Rev. Psychol. 36, 307–348. doi: 10.1146/annurev.ps.36.020185.001515

Wickens, C. D., and McCarley, J. S. (2008). Applied Attention Theory . London: CRC Press.

Wiener, E. L., and Attwood, D. A. (1968). Training for vigilance: combined cueing and knowledge of results. J. Appl. Psychol. 52, 474. doi: 10.1037/h0026444

Wilson, G. F., and Russell, C. A. (2003). Operator functional state classification using multiple psychophysiological features in an air traffic control task. Hum. Factors 45, 381–389. doi: 10.1518/hfes.45.3.381.27252

Wilson, G. F., and Russell, C. A. (2007). Performance enhancement in an uninhabited air vehicle task using psychophysiologically determined adaptive aiding. Hum. Factors 49, 1005–1018. doi: 10.1518/001872007X249875

Yang, L., Liu, Z., and He, B. (2010). EEG-fMRI reciprocal functional neuroimaging. Clin. Neurophysiol. 121, 1240–1250. doi: 10.1016/j.clinph.2010.02.153

Yue, G. H., Liu, J. Z., Siemionow, V., Ranganathan, V. K., Ng, T. C., and Sahgal, V. (2000). Brain activation during human finger extension and flexion movements. Brain Res. 856, 291–300. doi: 10.1016/S0006-8993(99)02385-9

Keywords : physical work parameters, physical fatigue, mental fatigue, vigilance, training, neuroadaptive systems

Citation: Mehta RK and Parasuraman R (2013) Neuroergonomics: a review of applications to physical and cognitive work. Front. Hum. Neurosci. 7 :889. doi: 10.3389/fnhum.2013.00889

Received: 23 August 2013; Accepted: 05 December 2013; Published online: 23 December 2013.

Reviewed by:

Copyright © 2013 Mehta and Parasuraman. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Ranjana K. Mehta, Department of Environmental and Occupational Health, School of Rural Public Healthy, Texas A&M University, Room 106, College Station, TX 77843-1266, USA e-mail: [email protected]

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

ERG-AI: enhancing occupational ergonomics with uncertainty-aware ML and LLM feedback

Open access
Published: 10 September 2024

Cite this article

You have full access to this open access article

Sagar Sen 1 na1 ,
Victor Gonzalez ORCID: orcid.org/0000-0002-4988-2425 2 na1 ,
Erik Johannes Husom 1 na1 ,
Simeon Tverdal 1 ,
Shukun Tokas 1 &
Svein O Tjøsvoll 3

269 Accesses

2 Altmetric

Explore all metrics

Workers, especially those involved in jobs requiring extended standing or repetitive movements, often face significant health challenges due to Musculoskeletal Disorders (MSDs). To mitigate MSD risks, enhancing workplace ergonomics is vital, which includes forecasting long-term employee postures, educating workers about related occupational health risks, and offering relevant recommendations. However, research gaps remain, such as the lack of a sustainable AI/ML pipeline that combines sensor-based, uncertainty-aware posture prediction with large language models for natural language communication of occupational health risks and recommendations. We introduce ERG-AI, a machine learning pipeline designed to predict extended worker postures using data from multiple wearable sensors. Alongside providing posture prediction and uncertainty estimates, ERG-AI also provides personalized health risk assessments and recommendations by generating prompts based on its performance and prompting Large Language Model (LLM) APIs, like GPT-4, to obtain user-friendly output. We used the Digital Worker Goldicare dataset to assess ERG-AI, which includes data from 114 home care workers who wore five tri-axial accelerometers in various bodily positions for a cumulative 2913 hours. The evaluation focused on the quality of posture prediction under uncertainty, energy consumption and carbon footprint of ERG-AI and the effectiveness of personalized recommendations rendered in easy-to-understand language.

Automated worker skill evaluation for improving productivity based on labeled LDA

Evaluation of a Real-Time Feedback Solution for Ergonomic Parameters Using Smart Sensors and User Centered Design

A novel platform to enable the future human-centered factory

Explore related subjects.

Artificial Intelligence

Avoid common mistakes on your manuscript.

1 Introduction

Human occupational ergonomics is important because it aims to design jobs, tasks, and work environments in a way that reduces the risk of musculoskeletal disorders (MSDs) and other types of injuries and health problems. Poor ergonomics can lead to musculoskeletal problems such as back pain, neck pain, carpal tunnel syndrome, tendinitis, and more [ 1 ]. It can also contribute to mental health problems, such as stress and burnout [ 2 ]. By designing interventions that reduce physical strain and promote healthy postures and work habits employers can improve well-being, reduce absenteeism, and create a more positive work culture. Sensors can be used in a variety of ways to measure and assess human occupational ergonomics. For instance, motion sensors can be used to track the movements of workers [ 3 ] and to assess the ergonomics of their workstation setup, force sensors can be used to measure the forces exerted on workers’ bodies [ 4 ] as they perform tasks, which can help to identify potential ergonomic risks, and temperature sensors can be used to measure the temperature of the work environment [ 5 ], which can affect the comfort and safety of workers [ 6 ]. By collecting multivariate data from these and other types of sensors, employers and occupational health professionals can get a better understanding of the ergonomic risks faced by workers and can take steps to mitigate those risks and improve the overall ergonomics of the work environment. Machine learning (ML) and artificial intelligence (AI) methods on wearable sensor data have been widely investigated to recognize how humans interaction with their work environment or their occupational ergonomics [ 6 , 7 ]. There is no doubt that there is a widespread trend of acquiring wearable sensor data in controlled settings and using AI/ML approaches to predict occupational ergonomics in the scientific community. Nevertheless, as far as we are aware, there is a scarcity of research regarding the engineering and implementation of AI/ML methodologies for occupational ergonomics in order to effectively integrate them into practical applications and facilitate human-friendly communication in natural language, with the ultimate aim of providing actionable insight to help mitigate musculoskeletal disorders (MSDs) among workers. Our motivation is driven by gaps in recent research at the intersection of AI and occupational health. Saadatnejad et al. [ 8 ] demonstrate the importance of modeling uncertainty in human pose forecasting, which enhances trustworthiness and reliability of the predictions. However, we see the need to go beyond pose forecasting by integrating a machine learning pipeline with Large Language Models (LLMs) to transform posture predictions and uncertainty estimates into actionable health risk assessments and recommendations. Furthermore, Saadatnejad et al. [ 8 ] utilize datasets like Human3.6M, AMASS, and 3DPW for evaluating pose forecasting models. None of these datasets are real-world datasets reflecting occupational health risks. To address this challenge we use a real-world dataset from the Digital Worker Goldicare dataset [ 9 ], which consists of several hours of data from home care workers. Furthermore, a recent comparative study [ 10 ], highlight that human experts outperform LLM-based ChatGPT in generating accurate and complete medical responses in occupational health. However, the study does not consider the use of real-world sensor data and uncertainty-aware posture detection to augment the responses from an LLM. This requires the development of comprehensive machine learning pipeline that can automate several steps of converting sensor data to posture over time followed by a natural language interpretation of it.

This study introduces ERG-AI, a machine learning pipeline designed to predict a sequence of postures from data collected over long-term observations via various wearable sensors. The pipeline leverages machine learning models’ performance metrics and predictive capabilities (e.g. uncertainty) to present occupational health risk assessments and improvements suggestions in a language that is easily understood by users. The ERG-AI pipeline incorporates a diverse suite of software modules for data input, feature extraction, data division, training, and inference employing AI/ML models such as deep neural networks and decision trees. It also acknowledges the presence of uncertainties in the workplace arising from measurement errors, limitations in sensing technology, user comfort and wearability, and calibration and maintenance needs. Given that it’s impractical to obtain independently and identically distributed data for every possible scenario, ERG-AI incorporates a ’dropout’ regularization scheme for epistemic uncertainty estimation, enabling random node elimination during training and inference stages. This facilitates ERG-AI to generate an ’uncertainty-aware confusion matrix’ (UCM) estimating the posture classification system’s accuracy and associated uncertainty. Ultimately, ERG-AI constructs a detailed prompt encapsulating the machine learning model’s posture sequence predictions and associated uncertainty estimates, which is processed by an API for a large language model (LLM) such as GPT-4 [ 11 ] and LLAMA-2 [ 12 ], to deliver a comprehensible occupational health risk assessment and user-specific recommendations.

We assessed the ERG-AI system using real-world data from the Digital Worker Goldicare dataset [ 9 ], which consists of 2913 hours of accelerometer data collected over 3.8 workdays from 114 home care workers. These workers, who include nurses, nursing assistants, occupational therapists, and others, wore five tri-axial accelerometers attached to various parts of their bodies. Previously, postures in this dataset were identified by applying vector mathematics to accelerometer data collected at a rate of 25 Hz within a measurement range of ±8 G. In designing a practical system for estimating occupational ergonomics risks, it’s crucial to consider factors such as battery usage and ease of use by workers. Consequently, we employed ERG-AI to predict posture based on down-sampled accelerometer data, reduced from 25Hz to 1Hz, in order to conserve battery life potentially on a low power device. We also studied the system’s performance as we incrementally reduced the number of physical sensors from five to just one. ERG-AI’s predictions were more accurate for certain postures (like lying, kneeling, sitting, and standing) using down-sampled data. However, it showed higher uncertainty when predicting actions like walking, running, climbing stairs, and other less common postures. As expected, we observed that reducing the number of accelerometers feeding into ERG-AI weakened its predictive capabilities. For example, when only one sensor was used on the arm, it could only reliably detect postures like lying, sitting, and to some extent standing and running. The system also generated higher uncertainty estimates for predictions when using fewer accelerometers. Taking into account the predictions for various postures and their uncertainty estimates over a period of time, ERG-AI generates a prompt to determine occupational health risks and recommendations. ERG-AI invokes an LLM API such as that of commercial GPT-4 [ 11 ] and open-source LLAMA-7B [ 12 ] to generate occupational health risks and recommendations. An ergonomics professional provided an analysis of the output to evaluate its validity and usefulness. We found that ERG-AI could provide meaningful and balanced explanations of occupational health risks, which are based on summary statistics of detected postures and the machine learning model’s performance and uncertainty. Nevertheless, there is a need for more specificity in the recommendations, where the knowledge of a worker’s age, gender and overall health can be beneficial. We also analyzed the differences between LLM outputs a commercial LLM GPT-4 and an open source LLM LLAMA-7B, which is much smaller in size and appropriate for portable devices. Lastly, we assessed ERG-AI in terms of energy consumption and carbon footprint for the training and evaluation phases to evangelize reporting our environmental footprint. We do not evaluate the environmental footprint of LLM inference in this article as it has been analyzed by other authors [ 13 ].

To summarize, our contributions include:

Comprehensive Machine Learning Pipeline for Occupational Ergonomics : ERG-AI incorporates a robust machine learning pipeline that handles data ingestion, preprocessing, model training, and deployment.The pipeline is designed to efficiently manage and process large datasets, ensuring accurate posture prediction and effective implementation in real-world scenarios, with an emphasis on energy efficiency and reduced carbon footprint. Furthermore, it leverages the DigitalWorker Goldicare dataset [ 9 ] for posture prediction which is a real-world dataset for occupational health of home care workers.

Uncertainty-aware Posture Prediction : ERG-AI predicts long-term worker postures using data from wearable sensors, incorporating uncertainty estimation to enhance prediction accuracy and reliability. The system generates an “uncertainty-aware confusion matrix” to evaluate the posture classification system’s accuracy and associated uncertainty.

Large Language Model-driven Insights for Occupational Ergonomics : ERG-AI leverages large language models (LLMs) like GPT-4 and LLAMA-2 to transform posture predictions and uncertainty estimates into comprehensible occupational health risk assessments and personalized recommendations for workers. The integration of LLMs ensures that the output is user-friendly and actionable, facilitating better health outcomes and ergonomic practices. To the best of knowledge, our article is the first work that combines posture prediction and large language models for generating occupational health risks and recommendations.

The rest of the paper is organized as follows. We present background work on sensor-driven occupational ergonomics, machine learning pipelines, uncertainty estimation, and LLMs in Section 2 . We then present related work on the use of AI/ML for occupational ergonomics in Section 2.6 and our machine learning pipeline ERG-AI in Section 3 . In Section 4 we evaluate the pipeline using sensor data acquired from care givers; and we finally conclude in Section 5 .

2 Background

In this section, we present background work on sensor-driven occupational ergonomics, data pipelines to pre-process sensor data and train machine learning models to predict posture, concept of uncertainty estimation to be incorporate in the data pipeline, large language models and dimensions of AI engineering that need to be considered for maintenance and deployment of machine learning models for occupational ergonomics.

2.1 Sensor-driven occupational ergonomics

Musculoskeletal disorders (MSDs) are injuries caused by stress on internal body parts such as muscles, nerves, tendons, joints, cartilage, and spinal discs during movement [ 14 , 15 ]. They impact many individuals across various occupations and industries, ranging from office work to manufacturing, construction, and healthcare. These disorders can lead to long-term disability and economic losses [ 18 ].

Musculoskeletal disorders caused by workplace activities are known as work-related musculoskeletal disorders (WMSDs). The high physical demands of certain jobs and the prevalence of WMSDs contribute significantly to the elevated sickness absence rates among workers. Numerous studies have utilized advancements in portable sensor technologies to accurately measure physical work demands [ 15 , 16 , 17 , 18 ].

The utilization of portable sensors for healthcare, wellbeing, and behavioral analysis to prevent WMSDs associated with awkward postures has been explored using both rule-based and machine learning models. These models aim to identify risks associated with specific tasks and the ergonomic design of tasks, tools, and workplaces to align physical jobs with workers’ natural capacities [ 19 , 20 , 21 ]. Developing accurate posture assessment tools requires collecting sufficient spatiotemporal work-related data. Traditional data collection methods, including self-reporting, manual observation, and sophisticated sensor networks, are time-consuming, intrusive, and require technical expertise that may not be readily available among workers and employers [ 9 , 18 ].

This research aims to design and test a methodology using an unobtrusive and automated data processing framework to classify body postures associated with the risk of developing MSDs, utilizing only wearable sensor technology (accelerometers) on workers during their activities. The activity classification output can identify ergonomic risk levels for each worker and major sources of ergonomic risks, aiding workers and decision-makers. The data used to validate the presented pipeline was pre-processed and labeled using a modified version of the custom-made software Acti4 (The National Research Centre for the Working Environment, Copenhagen, Denmark) [ 22 ]. Acti4 employs rule-based models to determine activity categories and postures, such as lying, sitting, standing, moving, slow walking, fast walking, running, cycling, stair-climbing, arm-elevation, forward trunk inclination, and kneeling, with high sensitivity and specificity [ 23 ].

2.2 Machine learning pipelines

A machine learning pipeline is a set of interconnected steps that are designed to transform raw data into a final model that can be deployed to make predictions on new data. In this paper we are using supervised learning, the branch of machine learning where a model is trained to create a function for mapping input data to expected outcome values. We employ a popular data pipeline framework called Data Version Control (DVC) [ 24 , 25 ] to implement ERG-AI. A DVC pipeline has several stages that are used to manage and version large datasets, machine learning models, and other data-intensive projects. We briefly describe the role of each stage in DVC in this section.

Data ingestion: In this stage, raw data is ingested into the data pipeline from a data repository such a file system, database, or an API. For instance, raw data from wearable sensors is acquired using a serial peripheral interface, universal asynchronous receiver-transmitter (UART), Wi-Fi, or Bluetooth by a mobile app and stored on a file system.

Data preprocessing: In this stage, a pipeline profiles, cleans and prepares raw data for training by machine learning algorithms. Profiling the data [ 26 ] provides information on the data quality and insights into the distributions of the various features. The data pipeline typically makes use of external libraries for profiling such as Pandas profiling [ 27 ] and Great Expectations [ 28 ] for specifying domain-specific assertions on data quality if need be. The profiling statistics can be used to clean the data, which involves removing missing or invalid values, and minimize using data of poor quality. After cleaning, pre-processing entails feature extraction from raw sensor data which is transforming raw sensor data to a set of relevant and robust features. A pipeline can use an external Python library for feature extraction such as TSFEL [ 29 ] that provides about 60 features extracted from time series data. Feature extraction reduces the amount of data that needs to be processed and analyzed, while retaining the essential information required for the task at hand. Raw sensor data is often complex, noisy, and high-dimensional, which can make it challenging to work with and interpret. After feature extraction, both raw data and features need to be normalized if they are to be used for machine learning. Normalizing sensor data is a process of scaling the data to fit within a predefined range. This is often necessary because sensor data may have a wide range of values, and some machine learning algorithms may be sensitive to the scale of the data. Normalizing involves transforming the data to have a zero mean and unit variance. A data pipeline typically employs external libraries such as Sci-kit Learn [ 30 ] that provides a number of off-the-shelf algorithms for scaling data (e.g., min-max scaler, standard scaler). Finally, preparing scaled data for training entails restructuring data in the form of input-output pairs required by machine learning algorithms. A window size on the input features is typically specified to represent the time horizon of sensor data used to make a prediction. Detecting posture is a classification problem where input data is used to predict several classes of postures. In many cases, prepared data may be imbalanced, meaning that one or more classes are underrepresented compared to others. This can cause the machine learning algorithm to be biased towards the majority class, resulting in poor performance for the minority class. Hence it is also necessary to balance the scaled data . This refers to the process of adjusting the distribution of data in a dataset to ensure that it contains an equal number of samples from each class. Here the pipeline may employ techniques such as oversampling, under-sampling and generation of synthetic data to balance scaled data. A recent review of techniques for balancing is presented by Susan et. al. [ 31 ]. DVC manages how data is versioned and stored during all the steps of the preprocessing stage ensuring that during multiple runs of the pipeline only relevant artifacts are update hence improving performance.

Model Training and Storage: In this stage, machine learning models are trained using the processed data. This stage may also include tasks such as hyperparameter tuning, model selection, and evaluation. Trained models are stored in a separate storage system, which can be local or remote. DVC manages how models are versioned during training phases and monitors the dependencies with the preprocessing stage. It ensures that training only occurs when there are updates to the data available in the processing phase. DVC can skip the execution of certain stages and instead fetch the correct output from the cache if they have already run in the same configuration.

Model Deployment: In this stage, the trained models are deployed to a production environment, where they can be used for inference or prediction. This may include tasks such as creating an API or integrating the model into an existing software system.

2.3 Uncertainty estimation

Posture prediction for occupational ergonomics using ma-chine learning may be affected by two broad categories of uncertainty: data uncertainty (also referred to as aleatoric uncertainty) and model uncertainty (also referred to as epistemic uncertainty) [ 32 ]. Data uncertainty refers to the inherent variability in data that can impact the accuracy of posture prediction in occupational ergonomics. It can be caused by errors in the measurement of input variables such as tri-axial accelerometer vectors in different joints of the body resulting in inaccurate posture predictions. The causes for measurement uncertainty can be calibration errors, sensitivity to temperature, mounting errors due to sensor not being properly aligned and secured, sensor drift over time due to temperature changes and mechanical stress, the ski-slope problem in high-frequency accelerometers, signal noise due to electromagnetic interference, and variations in sampling frequencies due to memory and power limitations. Data uncertainty can also be caused by errors and biases in the training data used to develop posture prediction models resulting in poor generalization performance and inaccurate posture predictions for new subjects or tasks. Data uncertainty can be due to both inter-subject and intra-subject variability in human posture and movement resulting in inaccurate posture predictions especially for tasks or postures that are not well-represented in the training data. Finally, data uncertainty can also occur due to environmental factors that can affect posture, such as changes in lighting, temperature, or work equipment. Model uncertainty stems from the selection of machine learning model structure and its parameters. Different neural network architectures (e.g. CNNs, FCNNs, RNNs, Transformer models) have different structures and different types and number of learning parameters. Therefore, various models predict posture differently and are a source of uncertainty.

Uncertainty estimation in our context is the process of quantifying the degree of uncertainty or error in posture prediction. It can help predict posture in human occupational ergonomics by providing a measure of confidence in the posture prediction models. In this paper, we use deep neural networks (DNNs) for posture prediction and estimating uncertainty in its prediction. There are different approaches to estimating uncertainty in the prediction of posture as presented in [ 33 ]. We present three most common approaches below:

Monte Carlo Dropout during inference: This method [ 34 , 35 ] involves randomly dropping out some neurons during the forward pass of a trained deep neural network to obtain a set of predictions. The variance of these predictions can then be used as an estimate of model uncertainty.

Bayesian Deep Learning: This approach involves training a probabilistic model (e.g. Bayesian CNN/RNN) [ 36 ] that provides a distribution over model parameters. For instance, Bayesian neural networks can be trained using a variation of Bayesian inference called stochastic gradient Markov chain Monte Carlo (SG-MCMC) [ 37 ]. This approach allows for the estimation of a posterior distribution over the model parameters given the training data, which can be used to make predictions and quantify uncertainties.

Deep Ensembles: This method involves training multiple deep neural networks [ 38 ] with different initializations, and averaging their predictions at inference time. It is important to ensure diversity in the ensembles. This can be achieved by using different architectures, regularization techniques, or hyperparameters for each network. The variance of the predictions by the different architectures can be used as an estimate of model uncertainty.

In this article, we use Monte Carlo dropout during inference to estimate posture prediction uncertainty. Dropout was originally a regularization technique used in neural networks to prevent over-fitting [ 39 ] and improve generalization. It involves randomly dropping out (ignoring) a percentage of the neurons in a neural network during training, which helps to prevent the network from becoming too complex and memorizing the training data rather than learning generalizable patterns. However, we use dropout during inference to estimate uncertainty in the prediction of posture [ 34 , 35 ]. Monte-Carlo (MC) dropout is one method where forward-pass for inference is performed multiple times on a DNN with dropout enabled. Each forward-pass will randomly drop a neuron and produce a different output. This results in a distribution of predictions can be used to estimate the uncertainty of the model. By generating a distribution of predictions, we can quantify both the mean prediction and the variability (uncertainty) of these predictions. Specifically, MC dropout helps estimate epistemic uncertainty , which arises due to the model’s lack of knowledge and can be reduced with more data. We quantify the uncertainty for classification problems such as posture prediction by computing the entropy $\textrm{H}$ of the softmax output from the neural network:

In this equation, $\textrm{H}(X)$ is the Shannon entropy of the random variable X representing the softmax output of the DNN predicting posture. The colon (:) means “is defined as.” The summation symbol $\sum _{x \in N}$ means to sum over all possible values x that X can take on, and p ( x ) is the probability of X taking on the value x . Finally, $\log p(x)$ is the logarithm of p ( x ) with base e (the natural logarithm). The Shannon entropy provides a measure of the uncertainty in the predictions: higher entropy indicates more uncertainty, while lower entropy indicates more confidence in the predictions. This entropy-based uncertainty measure is crucial for applications where knowing the confidence level of predictions can inform subsequent decision-making processes.

Monte Carlo dropout is a popular method for estimating uncertainty in deep neural networks that can be easier to implement and more computationally efficient than other methods such as Bayesian neural networks and deep ensembles. Bayesian neural networks [ 36 ] require a significant amount of computation to learn the posterior distribution of model parameters and make predictions using Monte Carlo sampling. Deep ensemble [ 38 ], on the other hand, requires training multiple neural networks independently and then averaging their predictions, which can be computationally expensive. In contrast, Monte Carlo dropout provides a simpler and faster way to estimate model uncertainty by randomly dropping out neurons during inference and averaging the predictions over multiple samples. This method requires only a small amount of additional computation during inference and can be easily integrated into existing models. Due to it’s computational efficiency and ease of implementation, we opted for using Monte Carlo dropout to estimate uncertainty in our experiments.

2.4 Uncertainty-aware confusion matrix

A confusion matrix [ 40 ] for posture detection shows the performance of a classification model in predicting the postures of a human worker based on sensor data. The confusion matrix summarizes the number of correct and incorrect predictions made by the model for each posture class, as well as the types of errors made. The confusion matrix is a powerful visualization to obtain an overview of the performance of a model. Nevertheless, a confusion matrix does not directly show uncertainty because it only provides information on the number of correct and incorrect predictions for each class. It does not give information on how confident the model is in its predictions or the degree of uncertainty associated with each prediction.

We introduce the concept of an uncertainty-aware confusion matrix that shows the uncertainty of a classification model in terms of a quantity that indicates uncertainty assigned to each predicted class. The uncertainty can be quantified using the standard deviation from the prediction probability for a class or entropy in the prediction class. The uncertainty-aware confusion matrix is similar to the traditional confusion matrix, but it also shows uncertainty assigned to each predicted class in colored boxes as shown in Fig. 5 . The uncertainties in our case are entropies computed based on the dropout method presented in Section 2.3 . Visually, higher uncertainty/entropy is represented by darker red color while a lighter red color indicates lower uncertainty/entropy.

In the context of posture prediction, if the wearable sensor system is designed to classify a person’s posture into three categories (standing, sitting, and lying down), the uncertainty-aware confusion matrix would contain the entropy in the prediction of each posture class prediction. The matrix can be used to obtain overview of metrics such as the expected true positive rate or expected false positive rate, which provide a measure of the overall accuracy of the posture classification system and its associated uncertainty. Using dropout inference in conjunction with the uncertainty-aware confusion matrix we aim to convey more trustworthy DNN outputs in real-world scenarios where data may not be independent and identically distributed (IID).

2.5 Large language models and prompt engineering

Large language models (LLMs) are a type of artificial intelligence model developed to understand and generate human-like text. They are built using a machine learning architecture known as Transformer [ 41 ], which enables the model to comprehend context across long pieces of text and generate coherent, contextual responses. In this article, we use LLMs to convey occupational health risks and recommendations based on posture predictions and their uncertainties. This entails transforming categorical and numerical information generated by ERG-AI to natural language that is easier for humans to comprehend. ERG-AI can invoke different LLMs via an API. We experiment with both a commercial LLM such as GPT-4 [ 11 ] and a small and portable open source LLM namely LLAMA-7B [ 12 ].

GPT-4 [ 11 ], an iteration of the Generative Pretrained Transformer (GPT) series by OpenAI, represents a significant advancement in LLMs. It has been trained on a broad corpus of Internet text, allowing it to generate human-like responses in a wide range of languages and styles [ 42 ]. The official number of parameters in GPT-4 has not been disclosed but rumours claim that is uses about 1.76 trillion parameters. We use a paid subscription to the gpt-4-32K model using OpenAI’s API to analyze and reply to inputs prompts. The number 32K refers to the maximum number of input tokens.

Privacy preservation and edge deployment of LLMs can be pivotal in advancing occupational health ergonomics. By processing data on local devices or near the data source, edge deployment minimizes the latency usually associated with cloud-based solutions such as OpenAI’s GPT-4 API, enabling real-time analysis and feedback crucial for monitoring and improving ergonomic factors in a workplace. Furthermore, it significantly enhances privacy preservation as sensitive information regarding an employee’s health and behaviors is processed locally, reducing the risks associated with data transmission and storage on remote servers. Open-source models like LLaMA 2 offer a platform for developing and sharing models trained on publicly available datasets, fostering a collaborative environment for innovation while ensuring transparency and accessibility [ 12 ]. LLAMA-7B, being the smallest in the LLaMA model range with 7 billion parameters, is a particularly good candidate for edge deployment on personal mobile devices due to its balance between model size and performance. Its relatively smaller size could allow for efficient deployment on resource-constrained devices, like mobile phones, enabling real-time ergonomic analysis and feedback directly on a worker’s device, enhancing both occupational health ergonomics and privacy preservation.

Prompt engineering [ 43 , 44 ] is a technique used to query LLMs. It involves carefully crafting the input prompts to elicit desired responses from the model. The aim is to guide the LLM’s response in a specific direction, enhance the output’s quality, or achieve a certain style or tone. For occupational health risk assessments, LLMs can analyze events indicating prolonged standing, and suggest mitigations. For instance, it could recommend periodic rest, ergonomic footwear, or use of sit-stand workstations based on a body of health and safety literature [ 45 , 46 ]. However, it should be used with expert oversight due to its limitations. LLMs can be instrumental in summarizing the performance of machine learning models that predict posture using sensor data. They can digest complex data, performance metrics, and statistical information such as uncertainty estimates, and produce comprehensible, clear summaries that can be easily understood by non-experts [ 44 ]. This helps bridge the gap between the highly technical world of machine learning and practical, real-world applications, such as occupational health risk assessment.

2.6 Related work

Efforts for preventing MSDs include the development of workplace ergonomics assessment methods and strategies. These are based on the use of ergonomics rules to monitor the frequency and duration of physically demanding movements and repetitive awkward postures. Common ergonomic rules for posture assessment such as “Rapid Upper Limb Assessment” (RULA) [ 47 ] and the Ovako Working Posture Analyzing System (OWAS) [ 48 ] are commonly implemented through self-reports and visual- and video-based observations and are thus subjected to high levels of inaccuracy and costs, as well as being time-consuming [ 49 ]. Ergonomic assessments like RULA and OWAS, though essential, suffer from subjectivity, sampling bias, and are time-consuming and costly. Observer presence may alter worker behavior, these methods lack real-time feedback, and may inadequately assess complex postures.

Compared to RULA and OWAS, using Machine Learning (ML) and Deep Neural Networks (DNN) with wearable sensors such as accelerometers and Inertial Measurement Units (IMUs) can provide more objective, accurate, and real-time posture assessments, eliminating observer bias and reducing manual labor. However, they require large, high-quality datasets and careful implementation. Recent research has demonstrated the potential of utilizing Machine Learning (ML) models to identify workers’ postures and activities through the analysis of motion data collected by wearable Inertial Measurement Units (IMUs) [ 50 , 51 , 52 ]. The studies primarily depended on traditional ML models that necessitate manual heuristic feature engineering. However, this method can introduce engineering bias and potentially overlook the valuable information present in sequential motion data. As a solution, several researchers have started to leverage Deep Neural Networks (DNNs) for automated feature engineering to address the issue of worker posture recognition [ 20 , 21 , 53 ]. This approach has proven to be highly effective, yielding a recognition accuracy rate of 94% for construction workers. Building upon these advancements, the current study explores the possibility of monitoring the risk of Musculoskeletal Disorders (MSDs) through the detection of postures using wearable accelerometers and DNN methods.

Estimating uncertainty in Deep Neural Networks (DNNs) used for posture detection, based on wearable sensor data, is vital for understanding model reliability and identifying less trustworthy or ambiguous predictions. Multiple studies have employed Bayesian neural networks to quantify this predictive uncertainty, evidencing their effectiveness [ 54 , 55 , 56 , 57 ]. However, Bayesian methods’ computational complexity and extended training durations, especially with large-scale or high-dimensional data, can limit their real-time usage. Monte Carlo methods, on the other hand, offer an alternate perspective, utilizing the principle of randomness [ 35 , 58 ]. Specifically, the Monte Carlo dropout method presents an efficient and direct approach for estimating model uncertainty in posture detection [ 59 , 60 ]. This technique involves randomly “dropping out” a portion of the neural network during training, creating multiple variations of the model. This helps to understand the model’s behavior under various data conditions, making it especially useful for posture detection where data can be highly variable. In our study, we apply the Monte Carlo dropout method to estimate model uncertainty. Recent work on uncertainty estimation for pose prediction include work by Saadatnejad et al. [ 8 ]. They introduce models that incorporate inherent data noise and uncertainty priors, featuring novel epistemic uncertainty quantification through deep clustering, and achieving significant accuracy improvements in pose forecasting with the release of the UnPOSed library for standard evaluation. In contrast, we present a comprehensive machine learning pipeline for posture prediction, utilizing an uncertainty-aware confusion matrix and integrating LLMs for user-friendly health assessments, evaluated with real-world sensor data. Saadatnejad’s work focuses on both aleatoric and epistemic uncertainties, validated with standard datasets, and aims at general pose forecasting, providing an open-source library for research with significant prediction accuracy improvements. Meanwhile, we emphasize practical application using Monte Carlo dropout for uncertainty estimation, integrate a practical pipeline with LLMs for occupational ergonomics and health risk assessments, evaluate energy consumption, carbon footprint, and LLM outputs, offering practical tools for occupational health ergonomics.

The application of LLMs such as GPT-4 within occupational health represents a novel research area to our knowledge. Recent work has explored the use of ChatGPT in occupational medicine. Padovan et. al. [ 10 ] evaluate ChatGPT’s accuracy in answering occupational medical questions compared to human experts, highlighting its limitations and need for human supervision. Oviedo et. al. [ 61 ] highlight the risks of using ChatGPT for safety advice, noting its tendency to provide oversimplified and sometimes erroneous advice, lack of transparency, keyword sensitivity, and emphasis on individual responsibility. Sridi et. al. [ 62 ] present ChatGPT’s potential in enhancing data analysis, virtual assistance, task automation, education, and multilingual support in occupational medicine. However, it also points out challenges like ethical concerns, confidentiality risks, inaccuracies, the need for expert validation, and the issue of AI hallucinations. In recent work, Farquhar et. al. [ 63 ] present an approach to detect hallucinations in LLMs like ChatGPT and Gemini using semantic entropy, which measures uncertainty in meaning rather than specific word sequences. This method outperformed naive entropy estimation and other baselines across various datasets and LLMs, improving detection of incorrect answers. By clustering answers based on semantic equivalence and calculating entropy, the approach helps avoid unreliable outputs. This generalizable and unsupervised method enhances LLM reliability, especially in critical fields like law and medicine, by refusing to answer high-entropy questions, thus ensuring safer and more trustworthy AI-generated content. Our work addresses the challenge of simplification and misinterpretation by the design of contextual prompts and establishes expert-guided summarization protocols to ensure nuanced and context-aware recommendations. LLMs have the potential to offer a rich, contextual comprehension for interpreting and articulating sensor data related to occupational health hazards. Specifically, ’instruction prompting’ employs definite commands or inquiries to navigate an LLM’s responses. In the case of summarizing detected postures from sensor data, the model is instructed to transform the raw or processed sensor readings into a summary that is easy to understand. Singhal et al. [ 64 ] showcase cutting-edge uses of LLMs for medical knowledge, where LLMs when tuned with instruction prompts, demonstrate reasonable performance but still lag behind actual clinicians. Our objective is to utilize instruction prompts with the statistics and sequences of detected postures to summarize occupational health risks and recommendations in language that is easy for humans to comprehend.

ERG-AI Pipeline

ERG-AI Training Sequence Diagram

Finally, it is also important to consider the impact of future disruptions, such as human enterprises incorporating AI-driven risks and recommendations for occupational health. The NIOSH study [ 65 ] explored how future disruptions could impact occupational safety and health (OSH). Using strategic foresight, researchers identified nine critical uncertainties and developed four scenarios: Trusted Partnerships, Multi-Polar World of OSH, Race to the Bottom, and Rugged Individualism and Trustworthy Government. Key challenges include data access, direct-to-worker communication, and misinformation management. Recommendations emphasize modernizing IT infrastructure, focusing on individualized OSH, and developing communication strategies. The study underscores the need for strategic planning, data management, and partnerships to enhance OSH readiness and resilience for future threats. Baldassarre et. al. [ 66 ] discuss the impact of generative AI and LLMs on occupational health practices, emphasizing the need for tailored ethical considerations. It highlights AI applications in workplace safety, HR tools, and healthcare, and outlines challenges such as data privacy, security, and misinformation risks. The European Parliament’s AI Act and WHO guidelines are referenced for regulatory frameworks. They advocate updating the ICOH Code of Ethics to incorporate transparency, human oversight, and data privacy. Recommendations include developing specific AI guidelines for occupational medicine to enhance worker safety and well-being, ensuring responsible AI integration in healthcare.

We present, ERG-AI, a machine learning pipeline as shown in Fig. 1 to generate occupational health assessments and recommendations based on raw wearable sensor data as input. The main actors in ERG-AI is an ML Engineer who configures the pipeline for either training/inference and the Worker who generates data using wearable sensors and receives an occupational health risk assessment with recommendations. We describe the training and inference lanes for the ERG-AI pipeline in the following subsections.

3.1 Training lane

The training lane as shown in Fig. 2 describes how the ERG-AI pipeline learns from labeled data and how the machine learning models are trained. We present the different stages of the training lane as follows:

Data Acquisition : The process starts with a Worker providing high-frequency multivariate sensor data. For instance, in our experiments we obtain data from five tri-axial accelerometers A domain expert labels this data using vector mathematics between joints to obtain postures such standing, walking, running, sitting and so on. This data is stored in the labeled data database ( LabelledDataDB ).

Data Pre-processing : The machine learning engineer ( ML Engineer ) then specifies the configuration for the ERG-AI pipeline. The pipeline extracts data from the LabelledDataDB as a CSV file and performs various preprocessing steps including data profiling for quality and data cleaning. The results of these preprocessing steps are stored in the file system used by ERG-AI. Feature engineering is performed on the cleaned data to extract statistical properties that can better inform the model about the underlying patterns in the data. Feature engineering extracts statistical properties, called features , from the raw input data that exhibit invariance to noise. Furthermore, the feature-based representations of time-series data [ 67 ] perform well in classifying tasks at a fraction of the computational cost of processing raw time-series data. The results of feature engineering are also stored in the file system. The ERG-AI pipeline splits the data into training, validation, and test sets and performs scaling/normalization on the data to ensure all variables are comparable. The data is then sequenced, meaning the sensor variables are divided into sequences of a certain length called the window size. The last stage of the preprocessing involves combining these sequences into a training data set and a test data set, which are then passed to the training and evaluation stages, respectively.

Training : The ERG-AI pipeline then trains and validates a machine learning model using the prepared training dataset and validation dataset, respectively. It uses the input time-series data (of chosen input window sizes ) data and output posture from the training set. The ML Engineer configures window size, learning parameters and ML model types (architectures). The ML Engineer may choose between various types of neural networks to train models: Fully-Connected Neural Networks (FCNN), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). These three network types can all work on multi-variate time series data. In addition, the ML Engineer may define the specific architecture of the neural network by specifying the exact number of layers and neurons of each layer. For CNNs, one may also specify whether or not to use maxpooling, and the maxpooling size. For RNNs, one can configure the unit type (Long Short Term Memory, Gated Recurrent Unit, etc). Before training, the pipeline sets apart a small portion of the training data (e.g., 20%) to use as a validation dataset . It automatically stops training the ML model if the prediction error of the validation dataset stops improving, preventing the over-fitting of the model to the training data. It saves the ML model for evaluation in the Model Database ( ModelDB ) for evaluation.

Evaluation : The model is evaluated using the unseen test dataset, and the performance metrics are stored in the ModelDB . By using an unseen test set, we treat the model as a black-box system, focusing solely on the inputs and outputs to assess its behavior without considering its internal mechanisms. This approach helps in evaluating how well the model generalizes to new data and minimizes bias due to hyper-parameter tuning during training. We compare the model output and the ground truth to determine the accuracy of the model’s predictions. To provide a visual representation of the model’s task performance, the pipeline generates plots of predictions on test data.

We use accuracy, F $_1$ -score, recall, and precision to evaluate the model performance. These metrics provide different aspects of the model’s effectiveness in classifying the data correctly. Accuracy is the proportion of true results (both true positives and true negatives) among the total number of cases examined. It is defined as:

where TP is the number of true positives, TN is the number of true negatives, FP is the number of false positives, and FN is the number of false negatives. Precision, also known as positive predictive value, is the ratio of correctly predicted positive observations to the total predicted positives. It is defined as:

This metric is important to detect when the cost of false positives is high. Recall, also known as sensitivity or true positive rate, is the ratio of correctly predicted positive observations to the all observations in actual class. It is defined as:

Recall is important when the cost of false negatives is high. The F $_1$ -score is the harmonic mean of precision and recall, providing a single metric that balances both concerns. It is particularly useful when the class distribution is imbalanced. The F $_1$ -score is defined as:

By using these metrics, we ensure a balanced evaluation of the model’s performance, taking into account both the accuracy of the predictions and the balance between precision and recall.

ERG-AI Inference Sequence Diagram

We also estimate model uncertainty in the predictions using Monte Carlo dropout (see Section 2.3 ) and obtain the uncertainty-aware confusion matrix (see Section 2.4 ) for all postures.

Deployment : The ERG-AI pipeline is then deployed as a Docker container model as a service with an inference API. The API embodies the trained, validated, and evaluated ML model as a service (e.g., Flask web service). The API is invoked using sub-sequences of data from input sensors and returns postures. Since the pipeline trains the ML model on input data features extracted from raw data and bounded by a scaling operation (e.g., values between 0 and 1), the model cannot always use the raw input sequences as they are. The feature engineering and inference operations require using ML libraries such as Sci-kit learn [ 30 ] and TensorFlow [ 68 ]. Therefore, parts of the pipeline used in inference, such as code to compute engineered features, scaler, and the ModelDB with all its dependencies (e.g., ML libraries), are packaged as a standalone container (e.g., docker). We can deploy the container on an edge device or on the cloud.

3.2 Inference lane

The inference lane describes how the ERG-AI pipeline uses the trained models to make predictions and how these predictions are used to generate occupational health risk assessments and recommendations. The sequence diagram of interactions for inference are shown in Fig. 3 .

Data Acquisition : The process starts with a Worker providing sensor data acquired over several days to ERG-AI Inference API . The API stores this new sensor data S in the sensor data database ( SensorDataDb ).

Configuration : The machine learning engineer ( MLEngineer ) then specifies the default ModelDB to be used via the API. The ML Engineer has elevated permission to call the inference API. He/she invokes the API with a token or credentials that grant rights to specify the model M that is retrieved by the ERG-AI Inference API . This invocation is typically done right after a new version of the API is released or the models in ModelDB have been updated in the training lane.

Inference : The ERG-AI Inference API uses model M to predict postures and uncertainties from the new sensor data S . The set of predictions P and their uncertainties U are stored in the file system. The API generates a simple statistical summary of postures and uncertainties in text T using (P,U) . For instance, the summary is the amount of time a worker is standing, sitting, walking, and so on over the period of time in S . One may also store the sequence of postures in the summary hoping that the change from one posture to another can reveal a potential occupational health risk that may not occur only due to prolonged sitting or standing.

Generating risks and recommendations : The Worker who seeks a summary of their occupational health risks and wants recommendations to improve invokes the ERG-AI Inference API with a short work description W that provides context. For instance, this can be a description of what a care worker does on a daily basis. The API combines W along with T summarizing model predictions and uncertainties to generate a prompt Pr for a large language model API such as GPT-4 or LLAMA-2. The prompt describes the work, a description of the model, the model’s uncertainties, a summary of its predicted postures, and an instruction to generate occupational health risks and recommendations. When the LLM API (e.g., GPT4 API Footnote 1 ) is invoked with Pr , it returns occupational health risks and recommendations based on the prompt and these are then provided back to the Worker .

3.3 Implementation of the pipeline

We have implemented the ERG-AI pipeline to generate occupational health risks and recommendations using Python and Data Version Control (DVC) Footnote 2 , a tool for structuring ML experiments and data versioning. Each pipeline stage is a Python program that takes input data and produces an output based on control parameters. DVC tracks the dependencies between input, output, and control parameters for each stage and stores the input and output in the cache for each pipeline execution. Therefore, DVC can automatically check if any pipeline stage has already run with the given input and control parameters. It can fetch the output from the cache instead of executing the pipeline again, reducing the computational resources needed for creating virtual sensors. The control parameters are in a configuration file separate from the source code. Thus, the user can explore various pipeline configurations (e.g., the type of machine learning model, window sizes of input and target sensors, splitting of data, and how to train ML models, selection of LLM such as GPT4 and LLAMA-2 to generate risks and recommendations) without knowing the implementation details of the pipeline.

We have integrated CodeCarbon [ 69 , 70 ], a framework for measuring the energy consumption and carbon footprint, into the ERG-AI tool. CodeCarbon offers the capability to monitor the energy usage associated with each stage of the pipeline. This not only provides valuable insights into the environmental impact of our machine learning system but also enables us to identify potential areas for energy optimization and sustainability improvements. The combination of CodeCarbon with Data Version Control (DVC) allows us to link the energy consumption data to specific pipeline executions, helping us understand the resource demands associated with different configurations and input parameters. This integration with CodeCarbon ensures that our research accounts for the ecological footprint of our AI-enhanced ergonomics system, aligning with our goal of creating a sustainable and efficient solution for the workforce.

The implementation of ERG-AI is open source and available on GitHub: https://github.com/SINTEF-9012/erg-ai .

4 Evaluation

In this section, we address the following Research Questions (RQ)s based on the Digital Worker Goldicare dataset:

RQ1. What is the performance in predicting human ergonomic posture from accelerometer data?

RQ2. Can uncertainty estimation gauge our trust in predictions of rare human ergonomic posture?

RQ3. Can reducing the number of sensors continue to give accurate predictions of human ergonomic postures?

RQ4. What is the energy usage and estimated carbon footprint of the various stages of the ERG-AI pipeline?

RQ5. Can LLM feedback on occupational ergonomics driven by uncertainty-aware ML output be actionable?

4.1 Subject of the evaluation

The Digital Worker Goldicare dataset [ 9 ] is made of data from home care workers with $\ge $ 50 employment (minimum 18.8 working hours a week), recruited from six of a total of 13 home care service units in Trondheim, the third largest city in Norway. Only workers that had direct contact with patients were included. All workers in these home care units were provided written and oral information about the research project and gave written consent before the study. Exclusion criteria were: (1) physical disability not allowing normal behavior, (2) office work, (3) bandage band aid and adhesives allergy, and (4) pregnancy. The study was conducted according to the Declaration of Helsinki and approved by the Regional Committees for Medical Research Ethics-Central Norway (No.: 64541). Five triaxial AX3 accelerometers (Axivity Ltd, Newcastle upon Tyne, UK) were mounted on the skin of the home care workers, using adhesive double-sided tape (3M; Witre, Halden, Norway) and secured with waterproof medical tape (Opsite Flexifix; Mediq, Oslo, Norway). They were worn 24 h per day for up to six consecutive workdays at a sampling frequency of 25 Hz and a measurement range of ±8 G. We down-sampled the data to 1 Hz to increase the practical usefulness of the system on low power devices. The accelerometers were attached to the following anatomical locations: (1) below the head of the fibula, on the proximal and lateral aspect of the calf, (2) on the distal, anterior and medial aspect of the femur (approximately 10 cm above the crest of the patella), (3) below the iliac crest of the hip, (4) the upper back approximately 5 cm to the side of the processus spinosus at the level of Th1-Th2 vertebrae, and (5) on the upper arm, approximately at the insertion of the deltoid muscle. The home care workers consisted of nurses, nursing assistants, learning disability nurses, and occupational therapists, having home care as their main employer, and worked an average of 38.5 h a week. For an average of 3.8 workdays, 2913 h of accelerometer data were recorded from 114 home care workers [ 9 ].

4.2 Results

RQ1: What is the performance in predicting human ergonomic posture from accelerometer data? To identify the most appropriate machine learning algorithm and architecture for our posture classification task, we combined manual experimentation with empirical validation, evaluating Fully-Connected Neural Networks (FCNN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN). We tested various configurations for all three network types by altering the number of hidden layers and nodes per layer. We used the ReLU activation function during all trials. For the CNNs, which are well-suited for capturing the temporal and spatial dependencies in accelerometer data, we experimented with different numbers of convolutional layers, kernel sizes, filters per layer, and the use of max-pooling layers. Recurrent Neural Networks (RNN), including Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), were also explored to capture temporal dependencies in the data. We tested various configurations by adjusting the number of recurrent layers and units per layer. Although LSTM and GRU units were effective in learning long-term dependencies, CNNs ultimately outperformed RNNs in terms of classification accuracy and computational efficiency. Throughout the experimentation phase, hyperparameter tuning was conducted manually, refining the learning rate, batch size, and the number of epochs, with a window size of 15 seconds used for each prediction to balance temporal context and computational efficiency. Our final CNN architecture, which yielded the best results, included four convolutional layers with 64 kernels each, followed by two fully-connected layers with 32 nodes each, using ReLU activation functions. The full list of configuration parameters is shown in Table 1 .

The overall performance scores of the models are summarized in Table 2 . Our findings demonstrate that the proposed ERG-AI framework exhibits high performance when predicting human ergonomic postures, even when operating with down-sampled data. Particularly noteworthy is the performance achieved when utilizing data from the complete set of sensors, including the arm, trunk, thigh, calf, and hip positions, where the overall accuracy reached 0.92 (see Table 3 ). Table 2 provides a breakdown of accuracy per posture class, showing high classification accuracy for postures such as lying, sitting and standing, while other classes exhibit lower accuracy rates across the various feature sets.

Accuracy for all classes on different sets of features

In Fig. 4 we present a visual overview of the accuracy of different posture classes across various sensor configurations. This plot offers a clearer and more immediate understanding of how the number and type of sensors influence the model’s accuracy for each posture class. It is evident that using the complete set of sensors (arm, trunk, thigh, calf, and hip) consistently results in the highest accuracy across almost all posture classes. This configuration is particularly effective for static postures such as lying, sitting, and standing, where the accuracy remains above 0.85. However, as the number of sensors is reduced, a noticeable decline in accuracy is observed, especially for dynamic activities like walking and running. For instance, when only the arm sensor is used, the accuracy for walking and running drops to 0.15 and 0.50, respectively, highlighting the importance of multiple sensor inputs for these activities. The “Stairs” class continues to exhibit the lowest accuracy across all sensor configurations, indicating the inherent complexity in predicting this particular activity. This is likely due to the diverse nature of stair-climbing movements, which may require a more sophisticated approach or additional sensors to capture accurately.

Overall, the results underscores the trade-offs between sensor configurations and classification accuracy. While the inclusion of more sensors generally enhances performance, it also increases the complexity and computational demands of the model. Therefore, a balanced approach is necessary to optimize both accuracy and efficiency for real-world applications.

The confusion matrix shown in Fig. 5 provides a visual representation of the model’s performance, highlighting the distribution of true positives, false positives, and misclassifications for each posture class. The results in Fig. 5 are from the model trained on feature set 1, corresponding to the first row of Table 2 . When examining the confusion matrix, we can observe that misclassifications of the “Walk” class predominantly shift into the “Other” class, whereas misclassifications of “Run” mostly transition to the “Walk” class. Stair-climbing is notably challenging to predict, with confusion occurring between “Walk,” “Other,” and the “Stairs” class. Notably, active postures, including “Walk,” “Run,” and “Stairs,” exhibit more confusion compared to static postures, indicating the increased difficulty in accurately predicting dynamic activities. The complexity of stair-climbing, which includes both ascending and descending stairs, adds an additional layer of intricacy to this class.

The “Other” class encompasses a broad range of activities that do not fall into the explicitly defined categories such as lying, sitting, standing, kneeling, or stair-climbing. This class is inherently challenging to classify accurately due to its diversity, potentially including activities that involve dynamic and unpredictable movements. The low accuracy score of 0.51 suggests that ERG-AI may face difficulties in distinguishing between these diverse activities accurately

Our results underscore the robustness of the ERG-AI system in classifying postures, especially those related to static positions. However, it is important to acknowledge the potential impact of higher sampling rates on the system’s performance. While our results with down-sampled data are promising, future investigations may explore whether higher sampling rates can further enhance the accuracy of posture classification. This exploration could provide valuable insights into the trade-offs between data granularity and computational efficiency.

RQ2: Can uncertainty estimation gauge our trust in predictions of rare human ergonomic posture? Uncertainty estimation involves generating a secondary output for each prediction. In our context, it entails quantifying the level of uncertainty or potential error in posture prediction. This quantified value serves as an estimate of uncertainty, essentially reflecting how confident the model is in its prediction. By providing this supplementary information, decision-makers can make more informed choices and gain greater confidence in their decisions. Additionally, uncertainty estimation enables the detection of potential anomalies or alterations in the surrounding context of the prediction of a posture.

We present the uncertainty-aware confusion matrix in Fig. 5 where the intensity of the red color indicates the level of uncertainty. This figure represents the results from the model trained on all available sensors (feature set 1). Darker the red the higher is the uncertainty in the prediction and vice versa. The model trained by ERG-AI in for our experiment demonstrates the low uncertainty when predicting Sitting, Running and Lying.The uncertainty is moderate for Kneeling, Standing, and Walking while there is high uncertainty in predicting Other and Stairs.It is also interesting to note that there is low uncertainty and low confusion between predictions of Sitting or Stairs and Lying and high uncertainty and moderate confusion when predicting Stairs and Other. Prediction of some postures have high uncertainty and high confusion such as Kneeling and Other as well as Stairs and Kneeling. The knowledge of uncertainty helps in targeting data acquisition to reduce confusion and and increase certainty in our prediction of postures. For instance, Stairs is a relatively rare event compared to other movements and hence it exhibits high uncertainty in its prediction. Balancing the dataset by acquiring more data involving climbing stairs can help reduce the uncertainty in predicting Stairs.

Uncertainty-aware confusion matrix for predictions on the test set using all sensors as input features

Accuracy and uncertainty for different feature sets. The bar plot shows the accuracy, while the lines represent the uncertainty levels for correct classifications and misclassifications

Additionally, Fig. 6 visualizes the relationship between prediction accuracy and average uncertainty across different feature sets. The bar plot illustrates the accuracy of posture prediction for each feature set, while the line plots indicate the uncertainty levels for correct classifications and misclassifications. As expected, models utilizing a more comprehensive set of sensors (e.g., Arm, trunk, thigh, calf, hip) exhibit higher accuracy and lower uncertainty. Conversely, models relying on fewer sensors show increased uncertainty and reduced accuracy. Understanding the interplay between accuracy and uncertainty enables more informed decisions regarding sensor deployment and data acquisition strategies.

A more detailed breakdown of the uncertainty is presented in Fig. 7 , showing the uncertainty estimation across all posture classes for every feature set. These results reinforces the conclusion that incorporating more sensors in general reduces uncertainty. For instance, the average uncertainty for correctly predicted “Sit” postures decreases from 8 with only the arm sensor to 0.26 when using the full set of sensors (arm, trunk, thigh, calf, and hip). This trend is consistent across most postures, highlighting the value of comprehensive sensor data in enhancing prediction confidence. However, certain postures such as “Stairs” and “Kneel” still exhibit high uncertainty even with the full sensor set, suggesting these activities are inherently more challenging to predict accurately. In contrast, the uncertainty for misclassifications remains significantly higher, indicating that the uncertainty estimation can be useful for identifying unreliable predictions and guiding further model improvements. This suggests that high uncertainty values can serve as a red flag for misclassifications, allowing for targeted interventions such as additional data collection or model refinement for the most problematic postures.

Uncertainty heatmap for all classes across the different feature sets

Uncertainty estimation provides a good indication of both good and poor performance for unforeseen data. For instance, as shown in Fig. 5 , good performance in correctly predicting a posture is correlated to low uncertainty and vice versa. For instance, all values with high accuracy in the diagonal vector have light red colors indicating low uncertainty. While, prediction of Stairs and Other have both low accuracy and high uncertainty. ERG-AI predicts posture on unforeseen wearable sensor data along with its uncertainty computed using Monte Carlo dropout. This uncertainty helps us know whether the prediction can be trusted and whether we need to improve the underlying model by acquiring more data for certain postures.

RQ3: Can reducing the number of sensors continue to give accurate predictions of human ergonomic postures? To investigate the impact of reducing the number of sensors on the accuracy and uncertainty in predicting ergonomic postures, we created four additional sets of features, each with progressively fewer sensors on different parts of the body. We trained five unique models for each set. Our goal was to understand how sensor quantity influences prediction accuracy, which is vital for system affordability, user comfort, and practical deployment. Table 3 presents data that illustrates the impact of reducing the number of sensors on the accuracy and uncertainty of predicting ergonomic postures. It compares five different sensor configurations, each with a varying number of sensors placed on different parts of the body: Arm, Trunk, Thigh, Calf, and Hip. Set 1 features a full array of sensors (Arm, Trunk, Thigh, Calf, Hip) for extensive coverage. Set 2 omits the hip sensor, while Set 3 additionally removes the calf. Set 4 limits to arm and trunk sensors, excluding lower body parts. Set 5 simplifies to an arm-only sensor setup.

The accuracy metric evaluates the effectiveness of each sensor configuration in predicting ergonomic postures, with higher values signifying better performance. For Sets 1 and 2, accuracy is comparably high at 0.92, suggesting minimal impact from the removal of the hip sensor. However, as sensors are further reduced in Sets 3 to 5, there’s a marked decline in accuracy, dropping to 0.91, 0.79, and 0.73 respectively. This trend highlights a direct correlation between the number of sensors and the ability to accurately predict postures, with fewer sensors leading to less effective posture prediction.

Uncertainty quantifies confidence in posture prediction, with lower values being more desirable. It’s divided into “Uncertainty (correct)” for accurately predicted cases, and “Uncertainty (misclassifications)” for errors. As sensor count decreases, uncertainty rises for both correct and incorrect predictions. This trend implies that with fewer sensors, the system’s confidence in its predictions diminishes, leading to more ambiguous outcomes, especially in misclassified instances. This increase in uncertainty with sensor reduction highlights the challenge of maintaining prediction reliability and clarity with a limited sensor setup.

The sensor placement’s relevance is evident in posture prediction accuracy. A comprehensive array (Set 1), with sensors on multiple body parts, achieves high accuracy and low uncertainty, emphasizing the need for diverse data points. Set 2’s exclusion of the hip sensor slightly increases uncertainty, indicating its role in prediction confidence, yet doesn’t heavily impact accuracy. However, further reductions in Sets 3 to 5, notably removing thigh and calf sensors, lead to significant drops in both accuracy and certainty. This underscores the critical role of lower body sensors, particularly on the thigh and calf, for precise and reliable ergonomic posture predictions.

The analysis of class-specific performance for ergonomic postures, based on data from various Tables ( 2 , 4 , 5 , 6 , 7 ), reveals significant variation in accuracy, precision, recall, F1-score, and uncertainty across different sensor configurations. Notably, the ’Lie’ posture maintains a high accuracy (0.99) and precision (0.99) across most sensor sets, with the lowest uncertainty (0.21) observed in the most comprehensive set (Set 1). In contrast, the ’Kneel’ posture experiences a dramatic accuracy drop to 0.00 and high uncertainty (8.88) as sensors are reduced. The ’Sit’ posture shows remarkable stability in accuracy (0.99) and consistent recall (around 0.98-0.99), indicating reliable detection. However, the ’Stand’ posture’s accuracy decreases significantly from 0.89 to 0.40, with a corresponding increase in uncertainty (up to 17.97) as the number of sensors is reduced. The ’Stairs’ posture presents the lowest accuracy (0.23) and high uncertainty (11.84), underscoring the challenges in its prediction. These findings highlight the high performance in detecting stable postures like ’Lie’ and ’Sit’ across sensor configurations, whereas complex movements like ’Kneel’ and ’Stairs’ pose substantial challenges, particularly with fewer sensors.

RQ4: What is the energy usage and estimated carbon footprint of the various stages of the ERG-AI pipeline? The ERG-AI pipeline includes functionality for measuring the energy usage of each stage of the pipeline, which enables monitoring and reporting of the resource consumption and carbon footprint of using the pipeline. By using the framework CodeCarbon [ 70 ], these metrics are automatically recorded each time the pipeline is run. By investigating the energy usage and carbon footprint of the various stages of the ERG-AI pipeline we can reveal important insights regarding the environmental impact of our machine learning model creation process.

Table 8 presents the duration, energy consumption, and carbon emissions for each pipeline stage. Every experiment was performed on computational infrastructure located in Norway, on a computer with 8 physical cores, 128GB CPU RAM and 2 NVIDIA A30 GPUs. The values for carbon emissions are based on that Norway have an average carbon intensity of 27.55 gCO $_2$ eq Footnote 3 (grams of CO $_2$ equivalents). The first two stages, “Profile” and “Clean”, were only performed once for the data set, since those stages were not affected by which features we chose to use as input variables to the model. The rest of the stages had to be rerun for each new feature set we explored, and we present here the mean duration, energy consumption and carbon emissions across the five different feature sets we used, in addition to the standard deviation for those five runs.

The “Train” stage, responsible for the actual model training, stands out as the most energy-intensive phase with an energy consumption of 0.842 kWh, and estimated emissions of 23.193 gCO $_2$ eq. This accounts for a substantial portion, $83\%,$ of both energy consumption and carbon emissions of the whole model creation process. This is as expected, since model training, especially for complex machine learning models like neural networks, demands significant computational resources. The “Clean” stage, responsible for the first part of the data preprocessing, and the “Evaluate” stage, which includes both model evaluation and uncertainty estimation, also contribute significantly to our carbon footprint, with an energy consumption of 1.352 gCO $_2$ eq and 3.035 gCO $_2$ eq, respectively. This is an order of magnitude lower than the “Train” stage, but an order of magnitude higher than the rest of the preprocessing stages. The “Clean” stage involves reading the full data set in order to clean up the data and remove unwanted parts, and employs heavy use of the Pandas framework [ 83 ], which is convenient to use, but relatively computationally expensive. The “Evaluate” stage involves employing Bayesian dropout on the neural network, which likely contributes to its higher emissions. Additionally, the creation of plots and visualizations within the “Evaluate” stage may explain its larger emissions compared to some of the preprocessing stages.

ERG-AI Generated Prompt

It is important to note that while the “Train” stage consumes the most energy, the preprocessing stages also make a notable contribution to our carbon emissions. As we consider the environmental impact of our machine learning pipeline, optimizing the energy usage and emissions in all stages, not just the training phase, becomes imperative. This optimization could include more energy-efficient algorithms, hardware, and data processing techniques to reduce our overall carbon footprint.

RQ5: Can LLM feedback on occupational ergonomics driven by uncertainty-aware ML output be actionable? ERG-AI generates a prompt summarizing the activity of a worker as shown in Fig. 8 . This prompt instructs an LLM API to analyze recorded activities of a healthcare worker in home care, including time spent in various postures and movements, along with a model’s prediction metrics like accuracy, precision, and uncertainty for each activity as computed by ERG-AI. The task is to assess ergonomic risks based on this data and provide five tailored recommendations to improve the worker’s occupational health and safety.

ERG-AI calls GPT4 to Generate Risks and Recommendations

The prompt in Fig. 8 is used to perform Retrieval-Augmented Generation (RAG) using a commercial LLM namely Open-AI’s GPT-4 and a smaller and portable open source LLM called LLAMA-7B. RAG is a technique in natural language processing that enhances text generation by integrating a retriever model to fetch relevant information from external sources, which is then used by a generator model to produce more informed and contextually accurate outputs. In our case, we fetch relevant information from the ERG-AI pipeline. The generator model is the LLM that produces more information given the context from ERG-AI.

Analysis of GPT-4’s output : The risk-assessment generated by Open-AI’s GPT-4 is shown in Fig. 9 . The ergonomic risk assessment customizes recommendations to their specific activities, like prolonged sitting and standing, offering actionable solutions such as sit-stand workstations and proper footwear. It also thoughtfully addresses uncertain ’Other’ activities by suggesting manual logging for greater clarity and emphasizes safety, especially in emergency scenarios. However, it falls short in detailing strategies for the ambiguous ’Other’ category and primarily concentrates on posture-related risks, potentially overlooking other ergonomic concerns like manual handling. Moreover, recommendations may face feasibility issues in diverse home care settings due to resource limitations and assumptions about environmental control. Additionally, the advice, while activity-specific, lacks personalization considering the worker’s unique physical conditions and preferences.

Analysis of LLAMA-7B’s output : The risk assessment using LLAMA-7B’s API for the same prompt in Fig. 8 is provided in Fig. 10 . The evaluation of ergonomic risks for a healthcare worker presents several pros and cons. Pros include a comprehensive set of recommendations addressing back pain, stress, and fatigue, practical solutions such as regular movement breaks and correct lifting techniques, a holistic focus on both physical and mental health, and the innovative use of technology like activity trackers. However, there are cons: some recommendations lack specific guidance on break duration and frequency, there are assumptions about the work environment that may not hold in all settings, and a reliance on the worker’s self-motivation and consistency in using tools like fitness trackers..

OpenAI’s GPT-4 vs. LLAMA-7B : Comparing GPT-4 and LLAMA-7B outputs reveals distinct approaches to ergonomic risk assessment. GPT-4 delves into specific activities like sitting and standing, offering detailed, actionable advice such as sit-stand workstations and proper footwear. It also acknowledges data accuracy issues, suggesting manual logging for unclear activities. Conversely, LLAMA-7B covers broader ergonomic risks with generalized advice on movement, lifting, and stress management, but lacks specificity. While GPT-4 proposes technology integration through manual logging, LLAMA-7B recommends using an activity tracker. GPT-4 validates its recommendations against specific activity data, emphasizing continuous monitoring. In contrast, LLAMA-7B’s conclusion is more general, without specific validation, offering a broader but less detailed perspective.

Feedback from Occupational Health Professional : LLM feedback on occupational ergonomics based on ML output is in line with general perspectives on healthy working conditions. Subject-specific feedback on identified musculoskeletal health risks are appropriate but still seem to lack specificity, specially related to each worker’s age, gender, overall health condition, specific role in the organization, all of which must be taken into consideration when providing recommendations for healthier working habits. An important aspect of occupational ergonomics is the design of tasks that fit the worker, rather than forcing a worker’s body into postures. Tnis is where more context information on the worker and their tasks will crucial in order to understand how the working activities and environments can be tailored to suit the needs of the worker in order to reduce the risk of musculoskeletal disorders.

ERG-AI calls open source LLAMA-7B to Generate Risks and Recommendations

4.3 Threats to validity

The ERG-AI study on enhancing occupational ergonomics with ML and LLMs faces several validity threats.

Internal validity : The ERG-AI system, designed to improve occupational ergonomics through machine learning and LLMs like GPT-4, encounters multiple internal validity challenges. The posture labels of the Digital Worker Goldicare dataset was identified using vector mathematics, which means that the ground truth of the dataset relies on the accuracy of that method. Confounding variables such as environmental factors or individual health issues could mislead conclusions about musculoskeletal disorders. The selection of home care workers from a specific region introduces selection bias, limiting the generalizability of results to other populations. History effects, including personal or professional events during the study, might independently affect the workers’ behaviors. Maturation effects, such as physical or mental changes over time, can also skew results, as can testing effects where familiarity with sensors alters natural movements. Instrumentation changes, involving modifications in sensors or data processes, threaten data consistency. Regression to the mean may falsely interpret natural score fluctuations as intervention impacts. Experimental mortality, or participant dropout, could bias outcomes if dropouts have different characteristics. Placebo effects might cause behavior changes due to belief in the intervention’s efficacy. Lastly, experimenter bias could influence data handling, affecting study conclusions.

External validity : The study’s external validity faces several threats. Its generalizability is limited as its findings, based on the Digital Worker Goldicare dataset, may not extend to varied worker types or work environments. The use of LLMs like GPT-4 and LLAMA-7B for ergonomic risk assessments as very specific choices and might not be universally relevant across all occupational settings. There is a need to customize LLMs for the purpose by fine-tuning them to mitigate this threat. Additionally, the study overlooks the environmental impact of LLM inference, an important aspect for the ERG-AI system’s broader applicability and sustainability. The study’s recommendations lack individual worker personalization, potentially narrowing its external applicability. Finally, the fast-paced evolution of AI and ML technologies could render the study’s findings less relevant over time, as new methods and data might offer different insights.

5 Conclusions and future work

5.1 conclusions.

ERG-AI represents a pivotal development in occupational ergonomics, amalgamating machine learning’s analytical robustness with the communicative proficiency of Large Language Models (LLMs) such as GPT-4 and LLAMA-7B. The system’s efficacy was validated using the Digital Worker Goldicare dataset, where it adeptly predicted various human ergonomic postures, especially static ones like sitting and lying, even under energy-efficient down-sampled data conditions. This aspect underscores ERG-AI’s practical utility in real-world scenarios. Our work notably investigated the impact of sensor reduction on the accuracy and uncertainty of posture prediction. Findings revealed that while basic postures were consistently identified with fewer sensors, complex movements such as kneeling and stair climbing presented significant prediction challenges. This emphasizes the crucial role of specific sensors, particularly thigh-based ones, and highlights a balance to be struck between system affordability and predictive accuracy. Incorporating uncertainty estimation into ERG-AI offered insights into the confidence level of posture predictions, thereby aiding in decision-making and pinpointing areas needing model enhancement. Nonetheless, this feature also contributed to an increase in the system’s energy use and carbon footprint, particularly during training and evaluation phases.

The application of LLMs to transform ERG-AI’s technical data into actionable ergonomic advice illustrated the potential for AI-driven occupational health risk assessments. Although GPT-4 and LLAMA-7B provided valuable insights, occupational health experts suggested the need for more personalized recommendations, considering individual worker attributes and specific job roles. This indicates an avenue for further refinement of LLMs to produce more tailored ergonomic advice.

To summarize, ERG-AI embodies the intersection of advanced AI techniques and occupational health, offering an innovative means to improve workplace ergonomics and worker welfare. Its proficiency in processing intricate sensor data and relaying findings in an accessible format renders it an instrumental tool in mitigating musculoskeletal disorders among workers. However, ongoing advancements and customizations are essential to optimize its effectiveness across varied occupational contexts.

5.2 Future work

Building on the foundational work of ERG-AI in occupational ergonomics, several future research directions can further enhance its utility and applicability:

GuardRails for Risk Assessment : Future research could focus on developing robust ’GuardRails’ [ 72 , 73 ] for Large Language Model (LLM) outputs in ergonomic risk assessments. These mechanisms would involve implementing checks and balances to ensure that the LLMs, such as GPT-4 and LLAMA-7B, provide not only accurate but also ethically and contextually appropriate recommendations. This may include developing algorithms to filter out biases, inaccuracies, or inappropriate suggestions from LLM outputs, ensuring that the advice given is safe, relevant, and practical for diverse occupational contexts. For instance, GuardRails can be used to constrain recommendations to be specific with regard to gender and age.

Portable Deployment : Advancing ERG-AI’s portability to different platforms using technologies like Intermediate Representation (IR) and frameworks such as Apache TVM [ 74 ] and Modular.AI’s Mojo language [ 75 ] could be another area of exploration. This would involve optimizing the pipeline for deployment on various hardware platforms, ensuring it is lightweight, efficient, and capable of running on devices with limited computational power. Such portability would facilitate the widespread adoption of ERG-AI, especially in remote or resource-constrained environments.

Privacy-preserving Federated Learning : Integrating privacy-preserving techniques like Federated Learning [ 76 ] into ERG-AI would enable the system to learn from decentralized data sources without compromising individual privacy. This approach allows the model to be trained across multiple devices or servers, ensuring that sensitive data does not leave its original location, thereby adhering to privacy regulations and enhancing user trust. Privacy-preserving federated learning (FL) has been explored with positive consequences for occupational health. The paper by Moe et al. [ 76 ] presents a novel approach to predict worker safety in construction environments using FL to train deep learning models on edge devices, enhancing safety management while maintaining data privacy. Prasad et al. [ 77 ] provide a comprehensive survey on FL in the Internet-of-Medical-Things (IoMT), addressing integration challenges and presenting a case study on blockchain-based FL for decentralized data analytics in healthcare. Additionally, Alahmadi et al. [ 78 ] introduce a privacy-preserved mental stress detection framework using IoMT and FL, demonstrating significant reductions in communication overhead and ensuring data security and efficiency.

Uncertainty-driven Continual Learning : Incorporating continual learning driven by uncertainty estimates [ 79 ] in ERG-AI could be a significant step forward. This approach would allow the system to continuously adapt and improve based on new data while being aware of its limitations. By focusing on areas with high uncertainty, the system could prioritize learning new or rare postures, enhancing its overall predictive accuracy and robustness.

Data Sovereignty and Data Spaces : Data sovereignty is the ability of an individual or organization to control how, when, and at what price its data is used across the value chain. A data space is like a digital ecosystem that brings together relevant data infrastructures and governance frameworks to facilitate data pooling and sharing [ 80 ]. Data space have the potential to significantly contribute to accelerating digital transformation within and across domains [ 80 ]. Data sovereignty provides the legal and ethical framework within which data spaces operate. Data spaces, such as International Data Spaces (IDS) represent the technological basis for trustworthy data exchange, supporting data sovereignty between businesses while complying with relevant standards, values, and regulations. Addressing data sovereignty [ 81 ] concerns in the context of occupational ergonomics is crucial. By focusing on data sovereignty and data spaces in the future development of ERG-AI, the pipeline can enhance its compliance with standards and legal requirements, encourage active participation of workers in studies and data sharing, and supports the use of insights for better workplace ergonomics, research, innovation and policy making. Insights gained from ERG-AI can be valuable for broader initiatives focusing on workplace health or the European Health Data Space.

Small Language Models : The use of Large Language Models requires significant computational resources. This entails not only a large energy consumption, but also often the need for using cloud services and thereby sharing potentially sensitive data. Smaller language models (SML) can make it easier to perform such tasks locally and on resource-constrained devices, reducing the carbon footprint and improving data control and privacy. LLAMA-7B, which we experimented with in this paper, is currently one of the smallest open source language models, but with the rapid development in the field of generative AI, we may see smaller, more capable models emerge, such as Orca 2 and Phi 2. Phi 2 outperformed LLAMA-2 in commonsense reasoning and language understanding [ 82 ]. It would be interesting to integrate such SLMs to process ERG-AI’s posture predictions and generate risk assessments and recommendations.

Feasibility studies : Future research will need to conduct an economic feasibility assessment to evaluate the cost-effectiveness of sensor deployment and data processing, considering both initial investments and long-term savings from improved ergonomic interventions. Social feasibility is underway, engaging with diverse stakeholders, including employees, occupational health experts, and industry representatives, to ensure the system’s acceptability and practicality in real-world settings. This involves addressing social ethics concerns, such as user privacy and data protection. Legal feasibility will be evaluated by reviewing relevant regulations and standards in occupational health and safety, data protection, and AI deployment. Ensuring compliance with laws such as the General Data Protection Regulation (GDPR) and other local data privacy laws will be paramount. Moreover, the legal implications of AI-driven ergonomic assessments will be considered, ensuring that the system offers recommendations while preserving human oversight and decision-making.

Limitations of this study : The study identified several limitations that warrant future exploration. Sensor dependency is a significant factor, with the accuracy of ERG-AI’s predictions being influenced by the number and placement of sensors; while basic postures were reliably identified with fewer sensors, complex movements required a more extensive network. To address this, future research could optimize sensor placement and explore alternative sensing technologies to reduce costs without sacrificing accuracy. Additionally, incorporating uncertainty estimation, though beneficial for decision-making, increased energy consumption, suggesting a need for more energy-efficient algorithms and hardware solutions. The generalizability of the model is another limitation, as validation was conducted using the Digital Worker Goldicare dataset, which may not represent all workplace scenarios. Building on the dataset to include a broader variety of tasks and environments would enhance the model’s applicability. Ethical concerns and data privacy requires continuous monitoring and updating to adapt to evolving standards and regulations. Future research will address these and other limitations by focusing on developing dynamic compliance frameworks that adjust to new legal and ethical guidelines in real-time. Lastly, ensuring unbiased and fair recommendations from the LLMs is crucial, we identified the need for further investigation into algorithms that can identify and mitigate biases, ensuring inclusive ergonomic solutions for all workers.

Data availability

Collected data and research protocols are available from the corresponding author on reasonable request.

https://platform.openai.com/docs/guides/gpt

Open-source Version Control System for Machine Learning Projects, https://dvc.org/

The carbon intensity for the current location is automatically fetched by CodeCarbon[ 70 ] from Our World in Data[ 71 ].

Holtermann A, Schnohr P, Nordestgaard BG, Marott JL (2021) The physical activity paradox in cardiovascular disease and all-cause mortality: the contemporary copenhagen general population study with 104 046 adults. Eur Heart J 42(15):1499–1511

Article Google Scholar

Cillekens B, Lang M, Van Mechelen W, Verhagen E, Huysmans MA, Holtermann A, Van Der Beek AJ, Coenen P (2020) How does occupational physical activity influence health? an umbrella review of 23 health outcomes across 158 observational studies. Br J Sports Med 54(24):1474–1481

Lim S, D’Souza C (2020) A narrative review on contemporary and emerging uses of inertial sensing in occupational ergonomics. Int J Ind Ergon 76:102937

Krugh M, Waldherr F, Gujarathi S, Kerner S, Mears L (2022) Wearable motion and force sensing to determine force exertion and task recognition for ergonomic analysis. In: IIE Annual Conference. Proc. pp 1–5. Institute of Industrial and Systems Engineers (IISE)

Sharma M, Suri N, Kant S (2022) Analyzing occupational heat stress using sensor-based monitoring: a wearable approach with environmental ergonomics perspective. Int J Environ Sci Technol 19(11):11421–11434

Donisi L, Cesarelli G, Pisani N, Ponsiglione AM, Ricciardi C, Capodaglio E (2022) Wearable sensors and artificial intelligence for physical ergonomics: A systematic review of literature. Diagn 12(12):3048

Fernandes V, Mendonça É, Palma ML, Nogueira M, Godina R, Gabriel AT (2022) Ergonomics and machine learning: Wearable sensors in the prevention of work-related musculoskeletal disorders. Occupational and Environmental Safety and Health IV 199–210

Saadatnejad S, Mirmohammadi M, Daghyani M, Saremi P, Benisi YZ, Alimohammadi A, Tehraninasab Z, Mordan T, Alahi A (2024) Toward reliable human pose forecasting with uncertainty. IEEE Robot Autom Lett

Tjøsvoll SO, Wiggen y, Gonzalez V, Seeberg TM, Elez Redzovic S, Frostad Liaset I, Holtermann A, Steiro Fimland M, (2022) Assessment of Physical Work Demands of Home Care Workers in Norway: An Observational Study Using Wearable Sensor Technology. Ann Work Exposures Health 052. https://doi.org/10.1093/annweh/wxac052

Padovan M, Cosci B, Petillo A, Nerli G, Porciatti F, Scarinci S, Carlucci F, Dell’Amico L, Meliani N, Necciari G et al (2024) Chatgpt in occupational medicine: a comparative study with human experts. Bioeng 11(1):57

Google Scholar

Waisberg E, Ong J, Masalkhi M, Kamran SA, Zaman N, Sarker P, Lee AG, Tavakkoli A (2023) Gpt-4: a new era of artificial intelligence in medicine. Ir J Med Sci (1971-):1–4

Roumeliotis KI, Tselikas ND, Nasiopoulos DK (2023) Llama 2: Early adopters’ utilization of meta’s new open-source pretrained model

Samsi S, Zhao D, McDonald J, Li B, Michaleas A, Jones M, Bergeron W, Kepner J, Tiwari D, Gadepally V (2023) From words to watts: Benchmarking the energy costs of large language model inference. In: 2023 IEEE High Performance Extreme Computing Conference (HPEC). pp 1–9. IEEE

Lunde LK, Koch M, Knardahl S, Wærsted M, Mathiassen SE, Forsman M, Holtermann A, Veiersted KB (2014) Musculoskeletal health and work ability in physically demanding occupations: study protocol for a prospective field study on construction and health care workers. BMC Public Health 14(1):1075. https://doi.org/10.1186/1471-2458-14-1075 . Accessed 03 March 2022

Nath ND, Chaspari T, Behzadan AH (2018) Automated ergonomic risk monitoring using body-mounted sensors and machine learning. Adv Eng Inform 38:514–526

Palm P, Gupta N, Forsman M, Skotte J, Nordquist T, Holtermann A (2018) Exposure to Upper Arm Elevation During Work Compared to Leisure Among 12 Different Occupations Measured with Triaxial Accelerometers. Annals of Work Exposures and Health 62(6):689–698. https://doi.org/10.1093/annweh/wxy037 . Accessed 03 March 2022

Vinstrup J, Jakobsen MD, Madeleine P, Andersen LL () Physical exposure during patient transfer and risk of back injury & low-back pain: prospective cohort study. BMC Musculoskelet Disord 21(1):715. https://doi.org/10.1186/s12891-020-03731-2 . Accessed 03 March 2022

Clays E, Hallman D, Oakman J, Holtermann A (2020) Objectively measured occupational physical activity in blue-collar workers: What is the role of job type, gender and psychosocial resources? Appl Ergon 82:102948. https://doi.org/10.1016/j.apergo.2019.102948 . Accessed 03 March 2022

Hendriksen PF, Korshøj M, Skotte J, Holtermann A (2020) Detection of kneeling and squatting during work using wireless triaxial accelerometers. Ergon 63(5):607–617. https://doi.org/10.1080/00140139.2020.1734668 . Accessed 03 March 2022

Zhao J, Obonyo E (2021) Applying incremental Deep Neural Networks-based posture recognition model for ergonomics risk assessment in construction. Adv Eng Inform 50:101374. https://doi.org/10.1016/j.aei.2021.101374 . Accessed 03 March 2022

Zhao J, Obonyo E (2020) Convolutional long short-term memory model for recognizing construction workers’ postures from wearable inertial measurement units. Adv Eng Inform 46:101177. https://doi.org/10.1016/j.aei.2020.101177 . Accessed 03 March 2022

Skotte J, Korshøj M, Kristiansen J, Hanisch C, Holtermann A (2014) Detection of physical activity types using triaxial accelerometers. J Phys Act Health 11(1):76–84

Korshøj M, Skotte JH, Christiansen CS, Mortensen P, Kristiansen J, Hanisch C, Ingebrigtsen J, Holtermann A (2014) Validity of the acti4 software using actigraph gt3x+ accelerometer for recording of arm and upper body inclination in simulated work tasks. Ergon 57(2):247–253

Kuprieiev R, Petrov D, Redzyński P, Pachhai S, Costa-Luis C, Schepanovski A, Rowlands P, Shcheklein I, Orpinel J, Santos F, Sharma A, Zhanibek, Gao, Taskaya B, Hodovic D, Grigorev A, Earl, Dash N, nik123, Vyshnya G, maykulkarni, Hora M, Vera, Mangal S, Baranowski W, Wolff C, Maslakov A, Khamutov A, Benoy K, Yoktan O (2021) DVC: Data Version Control - Git for Data & Models. Zenodo

iterative.ai (2022) Open-source Version Control System for Machine Learning Projects. https://dvc.org/ .(Visited in 2022)

Ruddle RA, Cheshire J, Fernstad SJ (2023) Tasks and visualizations used for data profiling: A survey and interview study. IEEE Trans Vis Comput Graph

Brugman S (2019) Pandas-profiling: exploratory data analysis for python. [Online]. https://github.com/pandas-profiling/pandas-profiling

Great Expectations (2022) https://greatexpectations.io/ . Visited in 2022

Barandas M, Folgado D, Fernandes L, Santos S, Abreu M, Bota P, Liu H, Schultz T, Gamboa H (2020) Tsfel: Time series feature extraction library. SoftwareX 11:100456

Nelli F (2023) Machine learning with scikit-learn. Python Data Analytics: With Pandas, NumPy, and Matplotlib. Springer. Berlin, Germany, pp 259–287

Chapter Google Scholar

Susan S, Kumar A (2021) The balancing trick: Optimized sampling of imbalanced datasets—a brief survey of the recent state of the art. Eng Rep 3(4):12298

Hüllermeier E, Waegeman W (2021) Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Mach Learn 110:457–506

Article MathSciNet Google Scholar

Gawlikowski J, Tassi CRN, Ali M, Lee J, Humt M, Feng J, Kruspe A, Triebel R, Jung P, Roscher R et al (2023) A survey of uncertainty in deep neural networks. Artif Intell Rev 56(Suppl 1):1513–1589

Ledda E, Fumera G, Roli F (2023) Dropout injection at test time for post hoc uncertainty quantification in neural networks. Inf Sci 645:119356

Gal Y, Ghahramani Z (2016) Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In: International Conference on Machine Learning. pp 1050–1059. PMLR

Wang H, Yeung DY (2020) A survey on bayesian deep learning. ACM computing surveys (csur) 53(5):1–37

Nemeth C, Fearnhead P (2021) Stochastic gradient markov chain monte carlo. J Am Stat Assoc 116(533):433–450

Rahaman R et al (2021) Uncertainty quantification and deep ensembles. Adv Neural Inf Proc Syst 34:20063–20075

Abdar M, Pourpanah F, Hussain S, Rezazadegan D, Liu L, Ghavamzadeh M, Fieguth P, Cao X, Khosravi A, Acharya UR et al (2021) A review of uncertainty quantification in deep learning: Techniques applications and challenges. Inf fusion 76:243–297

Liang J (2022) Confusion matrix: Machine learning. POGIL Activity Clearinghouse 3 (4)

Lin T, Wang Y, Liu X, Qiu X (2022) A survey of transformers. AI open 3:111–132

Radford A, Wu J, Child R, Luan D, Amodei D, Sutskever I et al (2019) Language models are unsupervised multitask learners. OpenAI blog 1(8):9

Ekin S (2023) Prompt engineering for chatgpt: A quick guide to techniques, tips, and best practices

Brown T, Mann B, Ryder N, Subbiah M, Kaplan JD, Dhariwal P, Neelakantan A, Shyam P, Sastry G, Askell A et al (2020) Language models are few-shot learners. Adv Neural Inf Proc Syst 33:1877–1901

Dunstan DW, Dogra S, Carter SE, Owen N (2021) Sit less and move more for cardiovascular health: emerging insights and opportunities. Nat Rev Cardiol 18(9):637–648

Wall R, Garcia G, Läubli T, Seibt R, Rieger MA, Martin B, Steinhilber B (2020) Physiological changes during prolonged standing and walking considering age, gender and standing work experience. Ergon 63(5):579–592

Wibowo AH, Mawadati A (2021) The analysis of employees’ work posture by using rapid entire body assessment (reba) and rapid upper limb assessment (rula). In: IOP Conference Series: Earth and Environmental Science, vol 704. IOP Publishing, p 012022

Gajšek B, Draghici A, Boatca ME, Gaureanu A, Robescu D (2022) Linking the use of ergonomics methods to workplace social sustainability: The ovako working posture assessment system and rapid entire body assessment method. Sustain 14(7):4301

Duncan S, Stewart T, Mackay L, Neville J, Narayanan A, Walker C, Berry S, Morton S (2018) Wear-time compliance with a dual-accelerometer system for capturing 24-h behavioural profiles in children and adults. Int J Environ Res Public Health 15(7):1296

Chen J, Qiu J, Ahn C (2017) Construction worker’s awkward posture recognition through supervised motion tensor decomposition. Autom Constr 77:67–81

Zhang X, Xie X, Ma L, Du X, Hu Q, Liu Y, Zhao J, Sun M (2020) Towards characterizing adversarial defects of deep learning software from the lens of uncertainty. In: 2020 IEEE/ACM 42nd International Conference on Software Engineering (ICSE). pp 739–751. IEEE

Ryu J, Seo J, Jebelli H, Lee S (2019) Automated action recognition using an accelerometer-embedded wristband-type activity tracker. J Constr Eng Manag 145(1):04018114

Kim K, Cho YK (2020) Effective inertial sensor quantity and locations on a body for deep learning-based worker’s motion recognition. Autom Constr 113:103126

Magris M, Iosifidis A (2023) Bayesian learning for neural networks: an algorithmic survey. Artif Intell Rev 56(10):11773–11823

Martinez-Hernandez U, Dehghani-Sanij AA (2019) Probabilistic identification of sit-to-stand and stand-to-sit with a wearable sensor. Pattern Recogn Lett 118:32–41

Zhao R, Wang K, Su H, Ji Q (2019) Bayesian graph convolution lstm for skeleton based action recognition. In: Proceedings of the IEEE/CVF International Conference on Compute Vision. pp 6882–6892

Martinez-Hernandez U, Dehghani-Sanij AA (2018) Adaptive bayesian inference system for recognition of walking activities and prediction of gait events using wearable sensors. Neural Netw 102:107–119

Blundell C, Cornebise J, Kavukcuoglu K, Wierstra D (2015) Weight uncertainty in neural network. In: International Conference on Machine Learning. pp 1613–1622. PMLR

Bangaru SS, Wang C, Busam SA, Aghazadeh F (2021) Ann-based automated scaffold builder activity recognition through wearable emg and imu sensors. Autom Constr 126:103653

Ogundokun RO, Maskeliūnas R, Misra S, Damasevicius R (2022) Hybrid inceptionv3-svm-based approach for human posture detection in health monitoring systems. Algorithm 15(11):410

Oviedo-Trespalacios O, Peden AE, Cole-Hunter T, Costantini A, Haghani M, Rod J, Kelly S, Torkamaan H, Tariq A, Newton JDA et al (2023) The risks of using chatgpt to obtain common safety-related information and advice. Saf Sci 167:106244

Sridi C, Brigui S (2023) The use of chatgpt in occupational medicine: opportunities and threats. Ann Occup Environ Med 35

Farquhar S, Kossen J, Kuhn L, Gal Y (2024) Detecting hallucinations in large language models using semantic entropy. Nat 630(8017):625–630

Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, Scales N, Tanwani A, Cole-Lewis H, Pfohl S et al (2023) Large language models encode clinical knowledge. Nat 1–9

Streit JM, Felknor SA, Edwards NT, Caruso DL, Howard J (2024) Preparing the occupational safety and health workforce for future disruptions. Am J Ind Med 67(1):55–72

Baldassarre A, Padovan M (2024) Regulatory and ethical considerations on artificial intelligence for occupational medicine. La Medicina del Lavoro 115 (2)

Lubba CH, Sethi SS, Knaute P, Schultz SR, Fulcher BD, Jones NS (2019) catch22: Canonical time-series characteristics. Data Min Knowl Discov 33(6):1821–1852

Pang B, Nijkamp E, Wu YN (2020) Deep learning with tensorflow: A review. J Educ Behav Stat 45(2):227–248

Lottick K, Susai S, Friedler SA, Wilson JP (2019) Energy usage reports: Environmental awareness as part of algorithmic accountability. arXiv preprint arXiv:1911.08354

Benoit C, Schmidt V. Goyal-Kamal M, Feld B, Lecourt J, SabAmine K, Léval M, Cruveiller A,Ouminasara FZ, Joshi A, Bogroff A, Inimaz, Saboni A, De Lavoreille H, Laskaris N, Abati E, LiamConnell, Blank D, Ziyao W,Catovic A, Michał S, Alencon JPW, MinervaBooks, SangamSwadiK, Bauer C, Hervé M (2023) Mlco2/codecarbon: CodeCarbon V2.3.1. https://doi.org/10.5281/ZENODO.8252426

Ritchie H, Rosado P (2020) Electricity mix. Our World in Data. https://ourworldindata.org/electricity-mix

Gasser U, Mayer-Schönberger V (2024) Guardrails: Guiding Human Decisions in the Age of AI, 1st edn. Princeton University Press Princeton, NJ

Book Google Scholar

Guardrails AI (2023) Guardrails: Adding Guardrails to Large Language Models. GitHub. Apache-2.0 License

Velasco-Montero D, Goossens B, Fernández-Berni J, Rodríguez-Vázquez Á, Philips W (2023) A pipelining-based heterogeneous scheduling and energy-throughput optimization scheme for cnns leveraging apache tvm. IEEE Access 11:35007–35021

(2023) Modular. https://www.modular.com/ . Accessed 01 Jan 2023

Moe SJS, Kim BW, Khan AN, Rongxu X, Tuan NA, Kim K, Kim DH (2023) Collaborative worker safety prediction mechanism using federated learning assisted edge intelligence in outdoor construction environment. IEEE Access

Prasad VK, Bhattacharya P, Maru D, Tanwar S, Verma A, Singh A, Tiwari AK, Sharma R, Alkhayyat A, Țurcanu FE et al (2022) Federated learning for the internet-of-medical-things: A survey. Math 11(1):151

Alahmadi A, Khan HA, Shafiq G, Ahmed J, Ali B, Javed MA, Khan MZ, Alsisi RH, Alahmadi AH (2024) A privacy-preserved iomt-based mental stress detection framework with federated learning. J Supercomput 80(8):10255–10274

Wang Q, Ji Z, Pang Y, Zhang Z (2024) Uncertainty-aware enhanced dark experience replay for continual learning. Appl Intell 1–16

(2020) Joinup: Data Spaces. https://joinup.ec.europa.eu/collection /semic-support-centre/data-spaces#1 . Accessed: 01 March 2024

Hummel P, Braun M, Tretter M, Dabrock P (2021) Data sovereignty: A review. Big Data & Society 8(1):2053951720982012

(2023) Microsoft: Phi-2: The surprising power of small language models. https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/ . Accessed 01 March 2024

The pandas development team (2020) pandas-dev/pandas: Pandas. Zenodo. https://doi.org/10.5281/zenodo.3509134

Download references

Acknowledgements

We would like to give a special thanks to Trondheim municipality, the home care units, unit leaders, and employees for outstanding collaboration and participation, despite the onset of the COVID-19 pandemic. This paper is supported by the European Union’s HORIZON Research and Innovation Programme under grant agreements No 101120657, project ENFIELD (European Lighthouse to Manifest Trustworthy and Green AI), and No. 101137207, project WAge (Healthier working environments for all ages), as well as by the Norwegian Research Council grant number 294762, project Digital Worker.

Open access funding provided by SINTEF.

Author information

These authors contributed equally to this work.

Authors and Affiliations

Department of Sustainable Communication Technologies, SINTEF Digital, Forskningsveien 1, 0373, Oslo, Norway

Sagar Sen, Erik Johannes Husom, Simeon Tverdal & Shukun Tokas

Department of Health Research, SINTEF Digital, Forskningsveien 1, 0373, Oslo, Norway

Victor Gonzalez

Department of Neuromedicine and Movement Science, Faculty of Medicine and Health Sciences, NTNU Norwegian University of Science and Technology, Edvard Griegs gate 8, N-7491, Trondheim, Norway

Svein O Tjøsvoll

You can also search for this author in PubMed Google Scholar

Contributions

SS: Contributed coordinating and conducting processing data, conducting analysis and data visualization, and writing the manuscript. VG: Contributed to designing, planning and organizing the study, data collection, pre-processing data, and writing of the manuscript. EJH: In charge of data analysis, contributed to writing of the manuscript. ST: Contributor of study data processing, writing, and feedback on the manuscript. ShT: Contributor writing the manuscript. SOT: In charge of data collection, and contributed to writing the manuscript.

Corresponding author

Correspondence to Victor Gonzalez .

Ethics declarations

Competing interests statement.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Ethical and informed consent for data used

All workers in these home care units were provided written and oral information about the research project and gave written consent before the study. Exclusion criteria were: (1) physical disability not allowing normal behavior, (2) office work, (3) bandage band aid and adhesives allergy, and (4) pregnancy. The study was conducted according to the Declaration of Helsinki and approved by the Regional Committees for Medical Research Ethics-Central Norway (No.: 64541).

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Sen, S., Gonzalez, V., Husom, E.J. et al. ERG-AI: enhancing occupational ergonomics with uncertainty-aware ML and LLM feedback. Appl Intell (2024). https://doi.org/10.1007/s10489-024-05796-1

Download citation

Accepted : 18 August 2024

Published : 10 September 2024

DOI : https://doi.org/10.1007/s10489-024-05796-1

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Human ergonomics
Machine learning pipeline
Large language models
Uncertainty estimation
Musculoskeletal disorders
Find a journal
Publish with us
Track your research

A .gov website belongs to an official government organization in the United States.

A lock ( ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Elements of Ergonomics
Upper Limb WMSD Consortium
View All Home

About Ergonomics and Work-Related Musculoskeletal Disorders

Ergonomics is the design of work tasks to best suit the capabilities of workers.
Ergonomics can help reduce or eliminate work-related musculoskeletal disorders (WMSDs) and other injuries and improve safety.
Ergonomics programs provide guidelines for reducing workplace risk factors, which can lead to reduced workplace injuries.

Workers in matching uniforms sorting cardboard boxes on a conveyer belt

Defining ergonomics

A man uses a crane to assist a lift in a manufacturing setting

In occupational health, ergonomics is the design of work tasks and job demands to fit the capabilities of the working population. The goal of ergonomics is to reduce and prevent musculoskeletal disorders caused by multiple factors. These include:

Physical (work tasks such as pushing, pulling, or lifting)
Psychosocial (mental wellbeing influenced by social factors)
Personal (age, sex, and body mass index)

In simple terms, the goal of ergonomics is to prevent discomfort and injuries caused by work.

NIOSH researchers recommend designing the work environment to fit employees' physical capabilities and limitations. Elements of the work environment can include:

This practice helps creates an ergonomically-sound work environment.

Terms to know

Using an understanding of ergonomics, employers, workers, and other parties can work together to design effective programs to prevent and minimize work-related discomfort and injuries. Here are some key terms to help you better understand ergonomics.

Musculoskeletal disorders (MSDs)

Disorders of the muscles, nerves, tendons, ligaments, joints, cartilage, or spinal discs that

Are caused by sudden or sustained force, vibration, repetitive motion, or awkward postures.
Are not the result of any instantaneous non-exertion event (e.g., slips, trips, or falls).
May be mild to intense and cause occasional to chronic pain.
Can be leading indicator for disability.

Work-related musculoskeletal disorders (WMSDs)

MSDs that can be attributed to the work environment and the performance of work.

Ergonomics program

A systematic process for identifying, analyzing, and controlling workplace risk factors, often for reducing MSDs.

Workplace risks

WMSDs can be caused or worsened by work conditions that require employees to:

Lift, push, pull, or carry many or irregularly shaped objects
Maintain awkward/unnatural postures
Withstand cold temperatures
Withstand vibrations from machinery and tools
Increase the intensity, frequency, and duration of activities
Endure psychosocial stressors

Learn more‎

Nioshtic-2 search.

See the NIOSHTIC-2 database search results on Ergonomics and Musculoskeletal Disorders . NIOSHTIC-2 is a database of occupational safety and health publications funded in whole or in part by NIOSH.

Health Hazard Evaluation reports

Search the NIOSH Health Hazard Evaluation (HHE) Reports for content related to ergonomics.

NIOSH Science Blog

Read the NIOSH Science Blog and share your comments.

The Occupational Safety and Health Act of 1970 established NIOSH as a research agency focused on the study of worker safety and health, and empowering employers and workers to create safe and healthy workplaces.

For Everyone

Work from Home and Disability Employment

There has been a dramatic rise in disability employment in the US since the pandemic, a pattern mirrored in other countries as well. A similar increase is not found for any other major gender, race, age or education demographic. At the same time, work from home has risen four-fold. This paper asks whether the two are causally related. Analyzing CPS and ACS microdata, we find the increase in disability employment is concentrated in occupations with high levels of working from home. Controlling for compositional changes and labor market tightness, we estimate that a 1 percentage point increase in work from home increases full-time employment by 1.1% for individuals with a physical disability. A back of the envelope calculation reveals that the post pandemic increase in working from home explains 80% of the rise in full-time employment. Wage data suggests that WFH increased the supply of workers with a disability, likely by reducing commuting costs and enabling better control of working conditions.

Funding to conduct the Survey of Working Arrangements and Attitudes was generously provided by the Templeton World Charity Foundation, Smith Richardson Foundation, Stanford University, Chicago Booth School of Business, Asociacion Mexicana de Cultura A.C., Stanford Institute for Human-Centered Artificial Intelligence, Toulouse Network for Information Technology, the MIT Mobility Initiative, and the Hoover Institution. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.

I worked for McKinsey and company as a management consultant from 2001-2002. I have not received any funding from them after that time.

I am part of the Toulouse Network for Information Technology, which carries out research on IT and productivity. From this network I receive an annual honorarium, which is funded by Microsoft.

I do occasional consulting on management practices for government and policy agencies, like the Canadian Government, the World Bank, the European Union, the British Government, and the European Bank of Reconstruction and Development.

I produced a report in 2008 for the World Economic Forum on management practices in private equity for which I received an honorarium.

I am a paid speaker at corporate events at which I discuss among other things working from home, management practices and policy uncertainty.

MARC RIS BibTeΧ

Download Citation Data

More from NBER

In addition to working papers , the NBER disseminates affiliates’ latest findings through a range of free periodicals — the NBER Reporter , the NBER Digest , the Bulletin on Retirement and Disability , the Bulletin on Health , and the Bulletin on Entrepreneurship — as well as online conference reports , video lectures , and interviews .

2024, 16th Annual Feldstein Lecture, Cecilia E. Rouse," Lessons for Economists from the Pandemic" cover slide

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Publications
Account settings

The PMC website is updating on October 15, 2024. Learn More or Try it out now .

Advanced Search
Journal List
Int J Environ Res Public Health

Intelligent Safety Ergonomics: A Cleaner Research Direction for Ergonomics in the Era of Big Data

Associated data.

Not applicable.

Safety ergonomics is an important branch of safety science and environmental engineering. As humans enter the era of big data, the development of information technology has brought new opportunities and challenges to the innovation, transformation, and upgrading of safety ergonomics, as the traditional safety ergonomics theory has gradually failed to adapt to the need for safe and clean production. Intelligent safety ergonomics (ISE) is regarded as a new direction for the development of safety ergonomics in the era of big data. Unfortunately, since ISE is an emerging concept, there is no research to clarify its basic problems, which leads to a lack of theoretical guidance for the research and practice of ISE. In order to solve the shortcomings of traditional safety ergonomics theories and methods, first of all, this paper answers the basic questions of ISE, including the basic concepts, characteristics, attributes, contents, and research objects. Then, practical application functions of ISE are systematically clarified. Finally, following the life cycle of the design, implementation, operation, and maintenance of the system, it ends with a discussion of the challenges and application prospects of ISE. The conclusion shows that ISE is a cleaner research direction for ergonomics in the era of big data, that it can deepen the understanding of humans, machines, and environment systems, and it can provide a new method for further research on safety and cleaner production. Overall, this paper not only helps safety researchers and practitioners to correctly understand the concept of intelligent safety ergonomics, but it will certainly inject energy and vitality into the development of safety ergonomics and cleaner production.

1. Introduction

Ergonomics (or human factors) is the scientific discipline concerned with the understanding of interactions among humans and other elements of a system, which is an important part of the ILO’s occupational safety and health activities [ 1 ]. Safety ergonomics, as an important research direction of safety science and ergonomics, mainly studies the relationship among humans, machines, and the environment from the perspective of safety and uses system analysis methods to achieve optimal matching among them. In recent years, with the development of modern communication technologies such as 5th generation (5G) technology, cloud computing, and the Internet of things (IoT), big data has become a focus of attention in academia, industry, and government departments around the world. Big data has prompted many experts and scholars to reexamine research methods, guided thinking, and the technological innovation of scientific research [ 2 ]. At present, a new generation of information technologies represented by big data is being increasingly adopted in various fields [ 3 , 4 , 5 ], such as in safety science [ 6 , 7 ] and cleaner production [ 8 , 9 ]. In this context, a new era of safety ergonomics represented by intelligent tools is rapidly approaching, and safety ergonomics is facing a new round of revolution.

The history of safety ergonomics has experienced a long development process. According to the development of system efficiency and relationships between humans and machines, the development of safety ergonomics can be summarized into three development stages: primitive ergonomics, ergonomics engineering, and modern safety ergonomics [ 10 ]. Since the beginning of human society, a simple interactive relationship has been formed between humans and machines [ 11 ]. In prehistoric times, to meet the survival conditions, humans enhanced their survivability by manufacturing tools. The initial and most primitive relationship between humans and machines was formed. After the industrial revolution at the end of the 18th century, labor-intensive and industrial-aggregating operations were advancing at a high speed, traditional manual manufacturing could no longer meet the development needs for productivity, and the requirements for tool reform became increasingly urgent. With the rapid advancement of mechanical physics research, early mechanized equipment was widely used in production and life. The development of electric power (the second industrial revolution) has greatly increased the number and types of production tools and equipment. To a certain extent, it also reduced the labor load of the personnel in the production system and increased the production efficiency of machines [ 12 ]. The First World War objectively promoted the development of ergonomics, and with the outbreak of the Second World War, the relationship between humans and machines gradually changed from "humans adapting to machines" to "machines adapting to humans," which also formed the theory of modern ergonomics engineering [ 13 ]. Since the 1950s, ergonomics has developed rapidly worldwide. However, safety issues in production have gradually emerged with the improvement of work efficiency [ 14 ]. In 1960, the concept of man-machine symbiosis was first proposed [ 15 ]. In the 1980s, the design of the interface and the method of interactive operations increasingly became a mainstream research direction [ 10 , 16 , 17 ]. In the 1990s, computer technology gradually matured, and computers slowly became a tool of control for manipulating complex production systems, and the process of mechatronics was accelerated [ 18 ]. Computer technology has changed the interaction forms and effects between traditional humans and machines, and it has accelerated the cognitive process of intelligent humans [ 19 , 20 , 21 ]. The human–machine relationship is also undergoing a new round of transformation, which provides a rare historical opportunity for the transformation and upgrading of safety ergonomics. The development history of ergonomics is shown in Figure 1 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g001.jpg

The developmental history of ergonomics. The development of safety ergonomics can be summarized into three development stages: primitive ergonomics, ergonomics engineering, and modern safety ergonomics.

With the advent of the Industry 4.0 era [ 22 , 23 ], the Internet of things and artificial intelligence (AI) have greatly changed the methods of working, producing [ 24 , 25 ], and thinking for humans. The complexity of the production system, as well as the interaction and collaboration between humans and machines is increasing [ 26 , 27 ], and the forms of human–machine collaboration to complete tasks has changed [ 28 ]; thus, the traditional accident causal model cannot satisfy the investigation and analysis requirements of complex system accidents and pollution [ 8 ], and the application of cleaner production faces huge challenges [ 29 ]. Therefore, the research content of safety ergonomics is no longer applicable to the analysis of new relationships among humans, machines, and the environment, and it is embodied in the following aspects: first of all, as intelligent communication technology overcomes the differences in time and space, the real-time operation of equipment and information integration becomes possible [ 30 ], and the traditional methods of interaction ways between humans and machines have been transformed into intelligent interaction. Secondly, the development and application of other disciplines [ 31 , 32 ] have injected more vitality into the research of safety ergonomics. At the same time, with the advent of the era of big data, information systems and various electronic computing devices are widely used in the practice of safety ergonomics. Whether it is the discipline of safety ergonomics itself or for the "humans, machines, and environment" elements included in safety ergonomics, big data technology has expanded its content and technology. Finally, compared with traditional classical disciplines, the disciplinary concept of safety ergonomics is lagging, and its theoretical foundation is still weak [ 33 ]. In particular, there are few cognitive analyses on the causes of systematic accidents, theoretical research and environmental pollution accidents.

Environmental research and public health are crucial for disaster control, environmental protection and human safety in production systems. The concept of environmental research and public health is being more and more widely used in different industries. However, there are still many problems in considering the safety and cleanliness of production systems using traditional safety ergonomics theories [ 34 ]. Therefore, based on the above considerations, in order to realize the coordination and unification among humans, machines and the environment in systems, and to promote the safe, correct, healthy, and efficient development of system operations, it is not only necessary to make full use of intelligent tools, but also to innovate the interactive modes among elements of systems. This is an inevitable trend to realize a new paradigm of safety ergonomics with intelligence as leading and core. In order to promote the development and practical research of safety ergonomics, we urgently need to answer the basic questions of safety ergonomics in the context of big data in a more detailed and systematic way. Therefore, in the second section, this paper proposes the basic concept of intelligent safety ergonomic (ISE), and analyzes the characteristics, attributes, contents and research objects of ISE through theoretical analysis. The third section puts forward the practical application functions of ISE. Finally, the future challenges and unsolved problems of ISE are analyzed in the final section.

2. Basic Questions of Intelligent Safety Ergonomics

2.1. what is it.

Safety ergonomics is a branch of safety science environmental engineering, and its scientific research and teaching tasks have been widely carried out worldwide [ 35 , 36 ]. However, in view of previous studies, safety ergonomics still does not have a universally recognized definition [ 37 ]. The authors proposed a new definition of safety ergonomics [ 38 ]: safety ergonomics is the use of knowledge in physiology, psychology, environmental science, artificial intelligence, and other disciplines, with the goal of safety and comfort, taking ergonomics as a condition, making humans, machines, and the environment coordinate with each other to meet people's growing needs for a better life and working environment, so as to achieve safety conditions. The research object of safety ergonomics is shown in Figure 2 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g002.jpg

Research object of safety ergonomics. The main research objects of safety ergonomics include humans, machines, and the environment, as well as the interactive interface between various elements.

With the continuous innovation and development of intelligent technology, the interconnected relationship between man—machine—environment is becoming increasingly complicated. In particular, the emergence of technologies such as intelligent human–computer interaction [ 39 , 40 ], biometric recognition [ 41 , 42 ], machine sensing, advance warning of disasters, and environmental adaptation [ 43 , 44 , 45 ] have greatly promoted safety ergonomics. The traditional definition is no longer applicable to the disciplinary system of safety ergonomics at this stage. Given the previous discussion, this section gives the following definition of ISE: intelligent safety ergonomics usually uses new-generation information technologies such as cloud computing, big data, 5G, and the Internet of things (IoT) as tools in combination with the knowledge of human factors, informatics, cybernetics, bionics, and environmental sciences to meet people’s growing demands for a better life and work as a starting point. It then focuses on the thorough integration of safety and environmental concepts into the design, implementation, operation, and maintenance of the human–machine–environment system throughout the life cycle, to achieve the intrinsic safety and cleaner production of the system as the goal, as to realize the self-adaptation, self-training, self-maintenance, self-learning, and self-optimization of the "human–machine–environment–information" system. The conceptual diagram of ISE is shown in Figure 3 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g003.jpg

The realization tools, basis of the subject, and point of departure, focus, target, and final goal of ISE.

The relevant theories in ISE are not unfamiliar. The new generation of information technology can be traced back to the field of computer engineering. The foundation of related disciplines can also be found in traditional human factors engineering. Some concepts of intelligent disciplines have also been proposed previously, i.e., intelligent mining [ 46 , 47 ], intelligent ergonomics [ 48 ], intelligent human factors engineering [ 49 , 50 ], etc. However, the concept of ISE is proposed for the first time. ISE considers the full life cycle of the system, which is completely different from the methods of traditional safety ergonomics.

With the development of information technology, especially in the context of big data, audio, video, and other massive data "blowout" growth [ 51 ], a wide range of data collection covers various fields of safety ergonomics research and human–machine–environment system data perception, and the processing technology is more intelligent; the internal data structure is also broader and more diverse. The differences between traditional safety ergonomics and ISE can be summarized in Figure 4 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g004.jpg

The difference between traditional safety ergonomics and ISE. The differences can be analyzed according to the tools, focus, purposes, objects, and knowledge.

2.2. What Are Its Characteristics?

Intelligence is the general term for intelligence and ability. At present, "intelligence" has been widely used in many fields [ 52 , 53 ], and it can be understood as a convenient tool to meet various needs in production and life, with the help of technologies such as computer networks and the IoT. Currently, breakthroughs have been made in the fields of machine vision, speech recognition technology, and equipment motion control, which are gradually integrated into other industries [ 25 ]. The development of intelligent industry and technology has promoted the diversification of human–computer interaction methods, and it has also improved the efficiency of these interactions. To a certain extent, it has also stimulated the development of safety ergonomics.

According to the basic concept of ISE, it can be considered that, ISE mainly coordinates the relationship among the elements of the system, integrating safety and the environmental concept into the whole life cycle (design, implementation, operation, and maintenance) of the system to the maximum extent. This is done to solve the problems of low safety performance, serious pollution problems, low degrees of intelligence, low self-adjustment and self-adaptive ability, the lack of guaranteed of intrinsic safety, and the difficulty in realizing safety control during the operation for system, thereby enriching the theoretical system of safety science. On a macro level, ISE has the following five main characteristics (advantages):

(1) The main research subject of ISE is still the "human–machine–environment" system. Realizing the efficient integration and optimal matching among the elements in a system is always the research focus in the field of ergonomics. However, with the development of information technology, the form of system interaction is gradually changing, and the previous normative interaction is also changing toward intelligent interaction. Especially with the development of sensor technology, "sensing" emerges in each interactive interface, and the information generation mechanism, transmission method, and interactive information performance characteristics among the elements of a system have undergone significant changes. Multi-channel information fusion interaction has gradually become a new form of interaction.

(2) The ultimate goal of ISE is to achieve intrinsic safety and cleaner production. It is difficult to achieve these goals in the construction of traditional safety ergonomics, but with the rapid development of information technology, intelligence has given us new technical methods. The goal of ISE is to carry out safety control and environmental protection measures, integrating safety and environmental concepts into the system throughout the life cycle of the system, so as to ensure the reliable operation of the system before the addressing critical safety and the clean state, thus achieving intrinsic safety and cleaner production. For example, in the life cycle of mine production, the degree of risk surrounding it can be judged by monitoring the subtle changes in rocks [ 54 ], employing the early implementation of safety control measures. It can be seen that with the use of the tools and theories of intelligent safety ergonomics, we can complete the safety measures that cannot be guaranteed by traditional technical means, and thus achieve intrinsic safety.

(3) The tool of ISE includes a broad new generation of information technology. In view of the existing intelligent manufacturing [ 55 , 56 ], intelligent visualization, and other means, the research on the integration of the new generation of information technology and safety ergonomics is still stuck in the theoretical integration of interdisciplinary theories, its in-depth integration and application need to be further developed.

(4) ISE is a discipline that combines theory and practice to guide the system to carry out safety and environmental control measures. First of all, through the theoretical analysis of ISE, it can guide management personnel to carry out safety practices, to further sublimate the essence of ISE. Secondly, ISE guides safety managers to carry out safety control practices, determine the safety state by analyzing the energy interaction among humans, machines, and environment, and implement safety and the environmental restraint measures through safety and environmental state feedback, thereby maintaining the dynamic balance of the system.

(5) ISE includes safety science, environmental science, and data science. First, the multi-element interactive fusion form widely existing in ISE is essentially the dissemination of digital information flow. Secondly, relying on the theory of ISE to solve practical problems, it is necessary to ensure that humans and devices have the capabilities of data collection, data analysis, and data processing, and then use data-driven approaches to develop related ISE content.

2.3. What Are Its Attributes?

Obviously, ISE is an applied branch of safety science and environmental engineering. Like other disciplines of safety science, ISE is the product of traditional safety ergonomics and intelligence. In addition, ISE intends to study the interaction and fusion method among the internal elements of a system, obtain the risk information of the system through intelligent perception, and then provide safety managers with a starting point and foothold for safety and environmental control methods, to achieve intrinsic safety and cleaner production. The specific attributes of ISE can be summarized in the following aspects [ 38 ]:

(1) ISE is a discipline that studies the coordination mechanism of the "human–machine–environment–information" system under the background of safety and the environment. With the advent of the information age, the rapid development of informational products has given birth to big data technologies based on massive amounts of data. In this context, the coordination mechanism among the traditional "human–machine–environment" systems has undergone fundamental changes, the communication among the various subsystems has changed, and the pattern and relevance of the identification data require a new safety ergonomics methodology.

(2) ISE is primarily affiliated with safety science and environmental engineering. First of all, its basic theory is still based on safety science and environmental theories. At the same time, realizing the intrinsic safety of the system is the ultimate goal of ISE. In addition, the theory of safety science involves studying the law of movement of safe things [ 57 ], and ISE is the study of the law of movement of the safety system with intelligent tools. It can be seen that ISE is still subordinate to the general and specific relationship between safety science and environmental engineering.

(3) ISE is a multi-disciplinary integration of safety and environmental science. The birth of ISE is an inevitable trend in the current information society, and it also involves the infiltration of other multidisciplinary disciplines into the field of safety and environmental science.

2.4. What Are Its Contents?

According to the research content, the content of ISE includes the following aspects [ 38 ].

(1) The basic theoretical level of ISE belongs to the upstream research content. The focus is on the research methods, analysis methods, design methods, basic principles, and subject attributes of ISE. Among them, the methodology of ISE focuses on the research on the scope, principles, theories, and methods of ISE to solve practical problems. It is usually based on ISE as the main body, summarizing and refining the research methods and paradigm systems that can solve practical problems. The basic theory of ISE refers to the research content and principles of other interdisciplinary subjects in the process of discipline construction. The discipline system of ISE mainly involves the basic issues of discipline construction. The upstream research content of ISE can be regarded as the scientific research on the subject. The basic theoretical level of ISE is shown in Figure 5 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g005.jpg

Basic theoretical level of ISE. Basic theory mainly includes three parts: methodology, basic theory, and subject system.

(2) ISE is an interdisciplinary subject, and the research on the basic level of its application is mainly conducted to extract scientific knowledge from the subjects it contains, and then to carry out research work on the basic problems of its field. The main research content of "intelligent humans," includes the decision–making skills, cognition, behavior, and physiology of humans, which can realize the functions of information perception [ 58 , 59 ], recognition, and prediction. The research purpose of "intelligent machines" includes intelligent perception, monitoring, detection, operation, and maintenance, etc. The research purpose of intelligent environmental is to realize the dynamic perception of the environment, and then to clarify the internal and external conditions of the production system. The specific content includes the static perception of the physical environment, the dynamic perception of the hidden danger environment, etc. [ 60 ]. The intelligent interaction can also be called the intelligent fusion technology of human–machine–environment elements, which refers to the situational cognitive function of intelligent humans and intelligent machines, and then to realize the semantic analysis of information parameters and intelligent processing of data information among each element. The application of the basic level of ISE is shown in Figure 6 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g006.jpg

Basic application level of ISE. Application of the basic level mainly includes intelligent interaction, intelligent humans, intelligent machines, and the intelligent environment.

(3) The purpose of any new subject is to solve practical problems. As far as ISE is concerned, the purpose of its practical application is to identify critical risk system information by identifying the failure and fault state of the system and to carry out safety control activities through intelligent monitoring, intelligent detection, and intelligent prediction. The practical application level of ISE contains many fields; specifically, it includes functional matching design for humans and machines, the optimal design of an intelligent interface, optimal intelligent design for work efficiency, and optimal intelligent design for intrinsic safety, etc. By carrying out research work at the practical application level, the risk level of the human–machine–environment system is converged, and then measures are provided for achieving the intrinsic safety and cleaner production of system. Practical application level of ISE as shown in Figure 7 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g007.jpg

Practical application level of ISE. Practical application level mainly includes function matching design, optimal design of interface, design for optimal work efficiency, and optimal design for intrinsically safe.

2.5. What Are Its Research Object?

Any production system always contains three parts: the human body, the mechanical equipment or body, and the environment in which they are located. Different from traditional safety ergonomics, ISE adds a new element-information based on the original research elements. The research object of ISE is shown in Figure 8 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g008.jpg

Research object of ISE. After receiving the information, humans use their brains to process original information to make safety decisions. Through interaction between humans and machines, intelligent control of the machines is realized, and results will be presented on the display.

Humans in traditional safety ergonomics are defined as natural persons who possess autonomy and creativity and who can explore and manipulate various devices, as well as make autonomous decisions. With the continuous acceleration of the intelligentization process, the labor cost in the production system has been greatly reduced, and the efficiency of production has also been fully improved. Some tedious tasks that can lead to fatigue characteristics in the human body are gradually being taken over by intelligent equipment. ISE proposes that the era of intelligence creates higher requirements for humans, and maintains that they are still the controllers of these intelligent machine. In particular, intelligent technology makes the production system more complicated, and human judgment is more susceptible to error. Therefore, humans are no longer individuals who deal with a single job in the traditional human–machine matching relationship, but "intelligent people" who can carry out intelligent decision-making tasks and control and break through the obstacles of human–machine interaction.

In ISE, a “machine” can be defined from a macro perspective as a mechanical device or energy carrier that can match human functions in a production system and can satisfy some specific functions. From the perspective of safety, the machine here is not only a specific mechanical device but also a macroscopic mechanism that can produce an unexpected release of energy. With the development of big data and the Internet of things technologies, intelligent devices continue to be popularized. These devices can realize functions such as computer processing, intelligent judgment, intelligent collaboration, and multi-channel interaction through sensors of the Internet of things, mobile networks, and other technologies. In turn, it can actively complete various tasks assigned by the production system.

The environment in ISE often includes the natural environment in which the humans and machines are located, such as humidity, temperature, and toxicity, which can be understood as the characteristics of the external conditions in the process of the human operation of the machine. In addition, it also includes the safety culture, safety atmosphere in the production system, and safety spirit. From a micro level, it can be understood as all environments that hinder or promote the integration of humans and machines. The psychological environment that affects human thinking is also included in this category.

In the phase involving the integration of humans, machines, and the environment, the timely regulation of the elements is essential to ensure the stability of the system and to achieve safety control. The interaction process of these three elements often requires a transmission carrier to facilitate communication, and information serves as such a function. It can be considered that the internal energy transmission and action mechanism of any abstract system are often transmitted by information, but it is different from traditional safety ergonomics, as the information in ISE is an interactive information flow created by external intelligent technology. The generation mechanism, action mechanism, transmission method, and transmission direction of traditional information flow are completely different. In addition, another reason why ISE is different from traditional safety ergonomics is the efficiency of information flow among the elements. The propagation mechanism of information affects the integration strategy of the human–machine–environment.

3. Practical Application Functions of ISE

From the definition and characteristics of ISE, the basic tasks of ISE can be obtained as follows: under the premise of giving the safety and environmental concept the leading role, the current emerging information technology tools are used to optimize the information flow dimension among humans, machines, and the environment to improve the intelligent coordination function of the system and to integrate it into all aspects of system analysis to achieve intrinsic safety and cleaner production. From the perspective of ISE practice, it pays more attention to the efficiency of human-computer interaction, the accuracy of system information transmission, and the interactive form of functional matching between elements. The coordination method and specific interpretation are shown in Figure 9 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g009.jpg

The function of intelligent coordination and its specific interpretation. By improving the intelligent coordination function of the system and integrating it into all aspects of system analysis, intrinsic safety and cleaner production can be realized.

3.1. Practical Application Functions of ISE

In order to further realize the above functions and understand the necessity of ISE, it is necessary to conduct an in-depth analysis of its functions. The so-called practical application function of ISE mainly refers to its safety and environmental control effect on human production and life. According to the focus of ISE, its practical application functions can be summed up as safety and environmental information perception, safety and environmentally intelligent monitoring, safety and environmentally intelligent detection, safety and environmentally intelligent prediction, and safety and environmentally intelligent operation and maintenance.

(1) Safety and environmental information perception. This is a basic function for the development of ISE application practice. Information is an important medium for ensuring functional matching among these three elements. ISE involves sensing technology; while people understand external information through physiological functions, the elements can also use intelligent sensing technology to complete the characterization and reasoning of external information through computing attributes to realize the interactive fusion among elements. For example, in mine production, we can determine the degree of risk by identifying potential risk areas and ensuring safe production in these areas [ 61 , 62 , 63 ].

(2) Safety and environmentally intelligent monitoring. This function includes relying on intelligent sensor technology to capture human physiological characteristics, health status, behavioral activities, language activities, location information, fault information, internal defects of mechanical devices, the information accessibility effect, the human–machine matching effect, and external environments such as the production environment temperature, wind speed, etc. According to the feedback results, the adaptive function adjustment in the intelligent coordination mode is used to cause the system to reach the best functional state. Existing studies have shown that it is feasible to use the Internet of things and cloud computing to establish a production early warning system for real-time monitoring, which also provides a technical guarantee for the feasibility of ISE to guide engineering practice [ 64 ]. Moreover, in the process of mine production, by using the current advanced positioning method (such as complex structure positioning [ 65 , 66 , 67 , 68 ]) to quickly determine the location of rock failure, it is also possible to determine the potential risk location in time according to different environments [ 69 , 70 , 71 ], thus realizing the purpose of early warning [ 72 , 73 ].

(3) Safety and environmentally intelligent detection. The safety status and stability of the system operation have a certain degree of connectivity, and there are hidden dangers in the operation process at any given time. To ensure the operation of the system, ISE requires regular intelligent detection activities, so as to determine hidden risks that may evolve into disasters, as well as the timely implementation of safety control activities aimed at controlling the potential risks of the system.

(4) Safety and environmentally intelligent prediction. Due to time constraints, the occurrence of accidents and disasters often lags. The monitoring function is continuous and occurs in real–time, and it does not have include a predictive function. However, as ISE can realize system inspection, detection, and data recording, "intelligent humans" can complete the evolutionary reasoning of the system through personalized cognition and then realize intelligent prediction.

(5) Safety and environmentally intelligent operation and maintenance: ISE emphasizes that systems can rely on big data technology and machine learning, collecting massive data (including logs, business data, system data, etc.) from a variety of data sources for real-time or offline analysis, thus enhance the ability of traditional operation and maintenance through initiative, humanization, and dynamic visualization to realize automated and intelligent decision–making.

3.2. Research Steps of Intelligent Safety Ergonomics

Intelligent safety ergonomics emphasizes providing intelligent perception functions to the system. The intelligent perception of the system mainly uses intelligent devices to monitor the relevant parameters or indicators that affect the unsafe behavior of humans, the unsafe state of machines, and the unstable changes of the environment during the operation of the production system, and analyzing, processing, and comparing the obtained results, to make dangerous or non-dangerous judgments regarding the system. This is a brand-new discipline whose research process, research, and application steps need to be explored continuously. Here, the general steps of the application of ISE in the practice of system safety control are presented.

(1) Analyze the characteristics of the researched system object, and select the relevant theories in ISE.

(2) Based on the cross-scientific theories contained in ISE, initially ascertain the common safety laws in the safety control process of the system.

(3) Propose a theoretical model of intelligent and safe ergonomics. ISE theoretical model is a theoretical framework that guides safety control technology to capture the critical safety state existing in the human–machine environmental system.

(4) Obtain the data sequence of the variable parameters of the critical safety state according to the safety information perception function of the intelligent system.

(5) Monitor changes in the critical safety state parameters, establish predictive models and causal chain analysis modules, and reveal the dynamic change law of the critical safety states.

(6) Combined with the dynamic change law of the critical safety state, discover the evolution process of the failure state for the system, use data dimensionality reduction methods to realize the dynamic visualization of the catastrophic evolution process of the system, and realize intelligent safety prediction.

(7) Combined with the intelligent cognitive decision-making mechanism, obtain the spatiotemporal strong mapping sequence of the node causality in the catastrophic evolution process, solve the catastrophic evolution link in reverse, analyze the final safety state and hidden danger position of the human–machine environment system, and carry out a scientific safety control strategy.

4. Prospects and Challenges of Intelligent Safety Ergonomics

The information processing technology and tools derived from the big data era have greatly enriched the development of safety science and the connotations of ergonomics, and they also provide new development ideas and methodology for safety ergonomics. First of all, ISE emphasizes the study of safety and environmental issues in the whole life cycle of the human–machine–environment system from the perspective of safety. Secondly, the advent of the big data era has overturned the traditional analysis methods of the systems, which has further expanded the research scope of these systems. Most importantly, the development of many new-generation artificial intelligence technologies has given birth to a broad, multi-dimensional system. With the help of the coordinated control function of the intelligent system, it can further provide a basis for solving the system function problems or the management of decision–making, reducing the risk of system disasters and environmental pollution to continuously improve and enhance the coordination and safety cleanliness of the system.

Mathematician David Hilbert said: “The branch of mathematics must be able to continuously produce new problems before it can be considered vigorous. Research fields that cannot ask questions are equivalent to a gradual death [ 74 ]”. With the rapid development of high-precision technology today, any discipline will have unresolved problems in its development process, and these problems also promote the progress and development of the discipline. ISE is the development and continuation of traditional safety ergonomics, and it is also an inevitable trend in the development of disciplines in the era of big data; thus, the future development of ISE still faces the following challenges.

Challenge 1: Intelligent perception and sensing technology of the human–machine– environment system.

The intelligent perception of the human–machine–environment system is the basis and key to realizing safe decision-making and control. Few studies regard it as a whole and reproduce it in the digital world, and most of the current sensing technologies usually realize intelligent perception through port and software analysis. As one of the core technologies in the era of big data, sensing technology is the key to the research of analog-digital signal processing, edge computing, and big data platform architecture. Although sensing technology has been applied to many fields, it is still difficult to build an intelligent system with integrated functions of sense, knowledge, and connection.

Challenge 2: Data acquisition and the information fusion mechanism of an intelligent system.

The advent of the era of big data has brought opportunities for the innovation and iteration of disciplines, as well as new tests for discipline practice. As big data technology guides a new round of disciplinary revolution, the ties between disciplines are moving closer and closer. However, it is a complicated, multi-dimensional, large-scale structure, which is difficult to clean, and which includes data and information island problems, heterogeneous multi-source multi-modal data fusion problems, and system mutation problems. At the same time, heterogeneous data and the fusion of multiple fields will have an impact on the safety control of the system. For example, in the process of mine production, by data collection and information fusion, the different risk levels of the production stage can be determined immediately, and the location of rock fracture and spatial evolution characteristics can be confirmed under complex stress conditions [ 75 ]. Therefore, the discussion of the data acquisition and information fusion mechanism of intelligent systems will become a major problem and challenge for ISE from cognitive theory to practical application.

Challenge 3: Design and development of a functional program of an intelligent and safe human–machine–environment system.

At present, intelligence has gradually become popular in visualization and data mining technology, but the degree of intelligence and automation is still not optimistic, especially for the dynamic quantitative characterization of a system disaster evolution mechanism, which is still difficult to achieve. In addition, the existing perception analysis technology still exhibits has the problems of a lack of information timeliness and a long time-consuming process. The design and development of the functional program for the system is still the focus of the future development of the discipline.

Challenge 4: High-speed storage and processing of intelligent system data.

Over the past 60 years, few scholars have focused on the use of human factors and ergonomic design to reduce system risks through effective interventions. Indeed, when the amount of data shows "blowout" growth, the efficiency of the useful information obtained from the data will be greatly reduced. Therefore, the high-speed storage and processing of data are also very important for the construction of the data architecture and the realization of the algorithm structure. It should be noted that due to the fact that big data has the characteristics of multiple sources, heterogeneity, and rapid change, the research on the computing paradigm of big data will also be the focus of future discussion because this is essential to the fine management of data quality and the collection of effective information.

Challenge 5: Implementation plan for the intrinsic safety and cleaner production of the full life cycle of the intelligent system.

In traditional ergonomics, since machines cannot realize the functions of self-diagnosis and self-feedback regarding their health status, errors will also occur regarding the fatigue operations of personnel, in this case, it is difficult to guarantee safe and clean production. With the development of big data tools, all of this has become possible. In addition, the purpose of building ISE is to explore the risk status of the human–machine–environment system and to achieve intrinsic safety and clean production through control methods. For example, in the production process of mines, microseismic monitoring technology, image recognition technology, artificial intelligence, and intelligent sensors can be used to develop intelligent mine driverless cars, thus realizing the intelligence of mining operations. Not only can it protect the safety of workers and equipment, it can also provide technical support for the safe and efficient recovery of resources [ 25 ]. Therefore, a reasonable implementation plan for intrinsic safety and cleaner production is also a challenge that ISE needs to face.

Challenge 6: Privacy protection and refined management of data.

Data is the fifth major production factor after land, labor, capital, and technology. Data security is a sensitive area in the process of data mining and utilization. In the era of big data, the statistical process of ergonomic safety data is more complete, and the scale of data continues to expand. Ensuring the confidentiality, integrity, and availability of the production system is also the data security solution pursued by ISE. Therefore, in the process of obtaining and using massive amounts of production data information, the protection of data privacy is both a technical issue and a social issue. For example, in order to ensure the sustainability of the mine production system, it is possible to analyze the multi-source heterogeneous data of the mine by establishing a human–machine–environment system evaluation method to realize safety control [ 76 , 77 , 78 , 79 ].

Challenge 7: Interaction and awareness insight mechanism for a human–machine–environment–system.

The key to the integration of system elements is to learn the behaviors of each aspect intelligently, realize awareness insights by finding common laws among fused data, and then construct a decision-making intent space to implement safety control. However, due to the complexity of the fusion mechanism, the interference of conscious behavior, and the ambiguity of state recognition, it is difficult for elements to realize each other's intentions. Therefore, it is very important to research a consciousness insight mechanism to realize the intelligentization of the system. In addition, the intelligent interaction process is subject to the existence of a wide range of information, and people's perceptions may be affected. The process of macroscopic intelligent mechanical manipulation may inhibit the understanding of information. Therefore, the technology to eliminate the visual noise regarding interaction interface is also a major challenge for ISE.

Challenge 8: Timeliness of information feedback regarding intelligent and safe ergonomics.

The limitations of big data technology, the complexity of multi-channel feature fusion for information transmission, and the uncertainty of human cognitive decision-making may cause a certain delay in the safety decision-making control for a system. Therefore, to improve the intelligentization process and efficiency of safety ergonomics, it is necessary to use artificial intelligence technology to speed up the design and development for new algorithms to ensure the timeliness of information feedback. Especially with the emergence of new imaging hardware and artificial intelligence chips, the design and development of intelligent algorithms for different chips and data acquisition devices is also a challenge.

Challenge 9: Operation and maintenance of the intelligent equipment of the system and the real-time control of the remote network.

On the one hand, intelligent operation and maintenance is a key technology to improve production efficiency, reduce safety management costs, and ensure safe production quality. It is also an important technical method to use intelligent technology to discover and solve abnormal operations of the system. However, due to the widespread existence of big data technology, wireless networks, and Ethernet networks coexist, and barriers to interconnection and intercommunication have increased, intelligent control and operation, as well as the maintenance of equipment groups are relatively difficult. On the other hand, it is difficult for the prior method to perform remote networking control of multiple device clusters through one control terminal, which also brings severe challenges to the intelligence of the machines. Dynamic decision-making component of intelligent equipment group operation and maintenance of the system is shown in Figure 10 .

An external file that holds a picture, illustration, etc.
Object name is ijerph-20-00423-g010.jpg

Dynamic decision-making component of intelligent equipment group operation and maintenance of the system. Dynamic decision-making components are based on the well-mined knowledge map of safe man-machine system operation and maintenance. It is an important means for realizing the intelligent control, operation, and maintenance of the equipment group.

Challenge 10: Intelligent interactive large-scale image perception and high-speed transformation technology.

The intelligent interaction of humans and machines is the focus of the research content for ISE. In order to realize the visualization of all data, it is necessary to seek common visualization technology and to ensure that intelligent machinery is highly sensitive to massive amounts of data, then presenting a large amount of safe information in different text forms. Therefore, the development of the intelligent interactive large-scale image perception and high-speed transformation technology of the "human–machine–environment" system is also the key content and major challenge requiring discussion.

5. Conclusions and Limitations

With the development of the Internet of Things and artificial intelligence technology, the way of life and production for humans are realizing the transition from informatization to intelligence. Systems are more complex, and the forms of interaction for the constituent elements are more diverse. Traditional safety ergonomics is no longer suitable for analyzing the elements in the new production relationship, and the iterative update of the theory is imminent. Intelligent safety ergonomics, as a new branch of safety science and environmental engineering, aims to thoroughly integrate safety and environmental concepts into the entire life cycle of the design, implementation, operation and maintenance of the human–machine–environment system; this is of great significance for achieving intrinsic safety and cleaner production. Of course, the research in this manuscript is not enough. The discipline construction of ISE still has a long way to go. Through analysis and research, we sorted out the following conclusions:

(1) Intelligent safety ergonomics, as an emerging notion created by the intersection of intelligent technology and safety ergonomics, is an inevitable trend for the iterative renewal of disciplines under the background of the era of big data, and it is also a cleaner research direction for ergonomics in the era of big data.

(2) With the aid of analyzing the definition of ISE, this manuscript answered the basic questions of ISE. At the same time, for further interpreting ISE, the practical application functions of ISE were systematically elaborated, and the future challenges and unsolved problems were analyzed.

(3) Although the concept of ISE has been proposed, its further research still faces many challenges. It must be made clear that this study is a preliminary exploratory study of ISE, and it does not provide comprehensive answers to its theoretical questions. It is still a lack of suitable case studies on how production systems can be used effectively. It can be seen from the challenges faced by ISE in the future that more extensive research on the scientific issues of the subject is needed, it is an important element that needs to be continued in future research to promote the development of theory and practice for safety science.

Acknowledgments

We are grateful for the financial support from National Key R&D Program of China (2021YFC2900500).

Funding Statement

This research was funded by National Key R&D Program of China, grant number 2021YFC2900500.

Author Contributions

Conceptualization, L.D. and J.W.; methodology, L.D.; resources, L.D.; writing—original draft preparation, J.W.; writing—review and editing, L.D.; visualization, J.W.; funding acquisition, L.D. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

“Not applicable” for studies not involving humans or animals.

Informed Consent Statement

Data availability statement, conflicts of interest.

The authors declare no conflict of interest.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

IMAGES

Holistic Ergonomics Infographic
5 Studies That Prove the Benefits of Ergonomics
Ergonomic research
Ergonomics work analysis method
5 Benefits of Ergonomics in the Workplace
Handbook of Research on Ergonomics and Product Design: 9781522552345

VIDEO

Ergonomics Research Techniques [Introduction Video]
Ergonomics Research Techniques week 3 assignment answer #nptel
Ergonomics Research Techniques
ERGONOMICS RESEARCH TECHNIQUES WEEK 2 ASSIGNMENT ANSWER #nptel
Ergonomics at Work: Comfort Meets Productivity
Design and Environmental Analysis Student Research

COMMENTS

Effect of an ergonomic intervention involving workstation adjustments on musculoskeletal pain in office workers—a randomized controlled clinical trial
Introduction. Computer usage has led to a progressive increase in both work and leisure time. Office workers spend long periods with static neck postures and exercising repetitive hand/wrist movements, often with awkward back postures 1, 2).These factors increase the chance of developing musculoskeletal pain, which can result in a reduced work performance, the need for sick leave, and even ...
(PDF) Literature Review on Ergonomics, Ergonomics Practices, and
Employing an explanatory research design anchored in the Theory of Ergonomics, the investigation aims to unveil the correlations between work stress, workload, decision-making, and communication ...
A systematic review of research on sitting and working furniture
The RULA is a method for assessing work posture through the use of a checklist that focuses on the upper body but does not exclude the lower body, and includes an assessment of the neck, the shoulders, trunk, wrists and arms in particular. ... There is a real paucity of research into the ergonomics of furniture used in everyday domestic life ...
Ergonomics
Ergonomics is the study of aligning a job's needs with the worker's and work environment's ability to provide the most efficient workspace possible while reducing the risk of injury. Historically, the primary goal of ergonomics has been reducing the rate of work-related musculoskeletal disorders (WMSD). The field assesses labor efficiency, quality, quantity, and comfort to maximize these ...
Ergonomics
The attention of ergonomics extends across work, leisure and other aspects of our daily lives. The journal Ergonomics is an international refereed publication, with a 60 year tradition of disseminating high quality research. Original submissions, both theoretical and applied, are invited from across the subject, including physical, cognitive ...
Ergonomics Advances in Office Work and Work-from-Home
Purchasing trends show that computer displays are increasing in size and new categories are offered, such as ultrawide (21:9 aspect ratio) and curved. Ergonomics research into these newer displays must keep up with purchasing trends (Gallagher et al., 2021). In this talk, the recommendations for multiple screens will be reviewed, and what is ...
Ergonomics: A bridge between fundamentals and applied research
Ergonomics is becoming a subject of applying fundamentals on anthropocentric dimensions for holistic welfare. The so-called conflict between Basic science and Applied research finds one of its edges in Ergonomics. Be it cutting-edge technology or frontiers of scientific innovation-all start from understanding basic scientific aptitude and ...
State of science: the future of work
Abstract. This article is concerned with scholarly ergonomics and human factors (E/HF) contributions to date to the field of research inquiry known as the 'future of work'. The review considers E/HF perspectives on how the nature of work is changing and what this means for the practice of E/HF and for human performance and wellbeing at work.
Office Ergonomics: A Review of Pertinent Research and Recent
Office ergonomics provides an arbitrary framework for integrating a large body of research that is relevant to the design of office work environments to optimize the health, safety, comfort, and effectiveness of their human occupants. In this chapter, I organize this vast literature by focusing on important empirical and practical developments ...
Bridging the gap between ergonomics methods research and practice
Ergonomics research and practice are underpinned by a suite of ergonomics methods that support the design or analysis of work and tools in relation to individuals, teams, organisations and even entire systems (Salmon, Walker, et al. Citation forthcoming; Stanton et al. Citation 2013).In a recent review of such methods, Stanton and colleagues identified well over 150 that cover important ...
Ergonomics in a rapidly changing world
The journal Ergonomics recently celebrated 60 years of publication and this special issue, with its heterogeneity of papers in terms of subject areas and context, reflects the maturity of ergonomics and human factors research. In this issue we have papers that address persistent topics in ergonomics/human factors research in a substantive way ...
Ergonomics at work
Ergonomics at work is a fundamental field of study that increasingly needs to gain space and. appreciation within organizations. It values th e well-being of its employees, creating, from its ...
Evidence-based ergonomics: a model and conceptual structure ...
The proposition on this study is to combine the theoretical approach for Sustainability Science, the Taxonomy of the Human Factors and Ergonomics (HFE) discipline and the framework for Evidence-Based Medicine in an attempt to be applied in Human Factors and Ergonomics. Applications of ontologies are known in the field of medical research and ...
Factors influencing ergonomists' use of observation-based risk
1. Introduction. The last few decades of ergonomics research concerning physical exposures and work-related musculoskeletal disorders (WRMSDs) has resulted in the development of a large number of different risk-assessment tools, e.g. self-administered questionnaires, interview protocols, observation-based risk-assessment tools (OBRATs) and direct (technical) measurement tools [].
Neuroergonomics: a review of applications to physical and cognitive work
Introduction. Neuroergonomics is defined as the study of the human brain in relation to performance at work and everyday settings (Parasuraman, 2003; Parasuraman and Rizzo, 2007).It integrates theories and principles from ergonomics, neuroscience, and human factors to provide valuable insights on brain function and behavior as encountered in natural settings (Parasuraman, 2011).
Principles, Methods and Examples of Ergonomics Research and Work Design
Helmut Strasser. 3139 Accesses. 4 Citations. The graphic term "ergonomics" as a combination of the Greek ergon, (1 Erg was the former term for work) and nomos (the Greek term for order or fundamental science) translated freely, has to do with the more bland work science, a science which is expected to contribute to the design of user ...
ERG-AI: enhancing occupational ergonomics with uncertainty ...
Building on the foundational work of ERG-AI in occupational ergonomics, several future research directions can further enhance its utility and applicability: GuardRails for Risk Assessment: Future research could focus on developing robust 'GuardRails' [72, 73] for Large Language Model (LLM) outputs in ergonomic risk assessments. These ...
Integrating Physical and Cognitive Ergonomics
The final article in this issue, by Garrett et al. (2016), provided a unique perspective on the business impact of integration of physical and cognitive ergonomics. Standing desks have gained recent attention as a workplace intervention to minimize sedentary behavior and improve physical activity among employees.
About Ergonomics and Work-Related Musculoskeletal Disorders
Ergonomics is the design of work tasks to best suit the capabilities of workers. Ergonomics can help reduce or eliminate work-related musculoskeletal disorders (WMSDs) and other injuries and improve safety. ... The Occupational Safety and Health Act of 1970 established NIOSH as a research agency focused on the study of worker safety and health ...
A study on the ergonomic assessment in the workplace
Sharan et al. [19] have done a study on the relationship between the ergonomic risk and the RULA assessment in. computer professions. A total n umber of 620 IT workers with the average age of 28. ...
Ergonomic interventions for preventing work‐related musculoskeletal
The remaining three studies assessed organisational ergonomic interventions of work breaks and work hours (Galinsky 2000; Galinsky 2007; McLean 2001). Although complete blinding for breaks was not possible in these studies, the use of a strict protocol for taking breaks by the use of either custom‐made electrical timers, or the 'Ergobreak ...
Ergonomics
Work-related MSDs can be prevented. Ergonomics --- fitting a job to a person --- helps lessen muscle fatigue, increases productivity and reduces the number and severity of work-related MSDs. Impact of MSDs in the Workplace. Work-related MSDs are among the most frequently reported causes of lost or restricted work time. A Process for Protecting ...
Work from Home and Disability Employment
Founded in 1920, the NBER is a private, non-profit, non-partisan organization dedicated to conducting economic research and to disseminating research findings among academics, public policy makers, and business professionals ... we estimate that a 1 percentage point increase in work from home increases full-time employment by 1.1% for ...
Occupational Ergonomics: A Special Domain for the Benefit of Workers
These findings have opened the door for more research on occupational health and ergonomics in the quest for health protection of vulnerable workers. ... (OSHA) definition, occupational ergonomics can be simply the study of work. More specifically, ergonomics is the science of designing the task to fit the worker, rather than physically forcing ...
How Trulicity Helps With Weight Loss
Caíque de Abreu/E+ via Getty Images . Living with diabetes comes with unique health challenges.Maintaining a healthy weight is one of them. Nearly 90% of U.S. adults with diabetes are considered overweight or obese.So naturally, many people with diabetes have questions about which diabetes medications can help you lose weight.. Trulicity (dulaglutide) is one potential option.
Intelligent Safety Ergonomics: A Cleaner Research Direction for
1. Introduction. Ergonomics (or human factors) is the scientific discipline concerned with the understanding of interactions among humans and other elements of a system, which is an important part of the ILO's occupational safety and health activities [].Safety ergonomics, as an important research direction of safety science and ergonomics, mainly studies the relationship among humans ...

Save citation to file

Add to My Bibliography

Evidence-based ergonomics: a model and conceptual structure proposal

Similar articles

LinkOut - more resources

REVIEW article

Introduction

Neuroergonomic Methods

Physical Work

Physical Work Parameters

Physical Fatigue

Cognitive Work

Mental Workload

Vigilance and Mental Fatigue

Training and Neuroadaptive Systems

Neuroergonomic Assessment of Concurrent Physical and Cognitive Work

Mobile Brain Imaging Considerations for Workload/Fatigue Assessments

Author Contributions

Conflict of Interest Statement

Acknowledgments

ERG-AI: enhancing occupational ergonomics with uncertainty-aware ML and LLM feedback

Cite this article

Similar content being viewed by others

Automated worker skill evaluation for improving productivity based on labeled LDA

Evaluation of a Real-Time Feedback Solution for Ergonomic Parameters Using Smart Sensors and User Centered Design

A novel platform to enable the future human-centered factory

1 Introduction

2 Background

2.1 Sensor-driven occupational ergonomics

2.2 Machine learning pipelines

2.3 Uncertainty estimation

2.4 Uncertainty-aware confusion matrix

2.5 Large language models and prompt engineering

2.6 Related work

3.1 Training lane

3.2 Inference lane

3.3 Implementation of the pipeline

4 Evaluation

4.1 Subject of the evaluation

4.2 Results

4.3 Threats to validity

5 Conclusions and future work

5.2 Future work

Data availability

Acknowledgements

Author information

Authors and Affiliations

Contributions

Corresponding author

Ethics declarations

Ethical and informed consent for data used

Additional information

Rights and permissions

About this article

Share this article

About Ergonomics and Work-Related Musculoskeletal Disorders

Defining ergonomics

Terms to know

Musculoskeletal disorders (MSDs)

Work-related musculoskeletal disorders (WMSDs)

Ergonomics program

Workplace risks

Learn more‎

Health Hazard Evaluation reports

NIOSH Science Blog

For Everyone

Work from Home and Disability Employment

More from NBER

Intelligent Safety Ergonomics: A Cleaner Research Direction for Ergonomics in the Era of Big Data

1. Introduction

2. Basic Questions of Intelligent Safety Ergonomics

2.2. What Are Its Characteristics?

2.3. What Are Its Attributes?

2.4. What Are Its Contents?

2.5. What Are Its Research Object?

3. Practical Application Functions of ISE

3.1. Practical Application Functions of ISE

3.2. Research Steps of Intelligent Safety Ergonomics

4. Prospects and Challenges of Intelligent Safety Ergonomics

5. Conclusions and Limitations