Wearable Sensors and Artificial Intelligence for Physical Ergonomics: A Systematic Review of Literature

Physical ergonomics has established itself as a valid strategy for monitoring potential disorders related, for example, to working activities. Recently, in the field of physical ergonomics, several studies have also shown potential for improvement in experimental methods of ergonomic analysis, through the combined use of artificial intelligence, and wearable sensors. In this regard, this review intends to provide a first account of the investigations carried out using these combined methods, considering the period up to 2021. The method that combines the information obtained on the worker through physical sensors (IMU, accelerometer, gyroscope, etc.) or biopotential sensors (EMG, EEG, EKG/ECG), with the analysis through artificial intelligence systems (machine learning or deep learning), offers interesting perspectives from both diagnostic, prognostic, and preventive points of view. In particular, the signals, obtained from wearable sensors for the recognition and categorization of the postural and biomechanical load of the worker, can be processed to formulate interesting algorithms for applications in the preventive field (especially with respect to musculoskeletal disorders), and with high statistical power. For Ergonomics, but also for Occupational Medicine, these applications improve the knowledge of the limits of the human organism, helping in the definition of sustainability thresholds, and in the ergonomic design of environments, tools, and work organization. The growth prospects for this research area are the refinement of the procedures for the detection and processing of signals; the expansion of the study to assisted working methods (assistive robots, exoskeletons), and to categories of workers suffering from pathologies or disabilities; as well as the development of risk assessment systems that exceed those currently used in ergonomics in precision and agility.


Introduction
Ergonomics deals with the design of work environments so that they are suitable for humans, and aims at the objectives of health and safety, and productivity at work [1]. Ergonomics as a discipline stands out for its systemic approach, design orientation, and the joint consideration of human well-being and performance [2].
Physical ergonomics is concerned with human anatomical, anthropometric, physiological and biomechanical characteristics as they relate to physical activity. Relevant topics include working postures, materials handling, repetitive movements, work-related musculoskeletal disorders (WMSDs), workplace layout, physical safety, and health [3].
• Physical ergonomics related to physical activity concerning human anatomical characteristics; • Cognitive ergonomics related to mental processes; • Organizational ergonomics related to optimization of socio-technical systems.
To the best of the authors' knowledge, no systematic reviews consider the potential combined use of wearable devices and AI algorithms in physical ergonomics applications. Some reviews have focused on the potential use of wearable devices in ergonomics [12,[23][24][25], while others have focused on the role of ML in the prevention of WMSDs [17,26]. This systematic review aims to fill this gap in the literature, considering the growing use of wearable devices and AI in medicine, and particularly in occupational medicine.

Research Strategy
The systematic review is a method of selecting, evaluating, and summarizing studies based on a specific topic [27]. Our systematic review is presented according to the Preferred Reported Item for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines [28].

Search Methodology and Study Selection
The literature search was conducted on Scopus and PubMed databases, and it was limited to English documents. Each database was queried using the following keyword structure: ("wearable" OR "sensors") AND ("ergonomics" OR "occupational medicine" OR "occupational health") AND ("AI" OR "ML").
In order to simplify our research, the exclusion criteria were: • Conference reviews, reviews, book chapters and erratum; • Papers not available; • Papers duplicated.
Concerning the screening by title, abstract, and full text, the following exclusion criteria were defined: • Papers proposing human-machine interface solutions without wearable devices, and not explicitly related to occupational medicine (e.g., touchless control interface in an underwater simulation environment [29]); • Papers proposing wearable devices for cognitive ergonomics (e.g., [30]); • Papers proposing only a wearable device solution without AI (e.g., [31]); • Papers proposing wearable devices for other purposes (e.g., rehabilitation [32]).
Documents were screened evaluating, firstly, title and abstract contents and, in case the documents did not meet the inclusion criteria, secondly the full text. Figure 1 shows the PRISMA workflow, and the number of documents included in this systematic review.

Research Strategy
The systematic review is a method of selecting, evaluating, and summarizing studies based on a specific topic [27]. Our systematic review is presented according to the Preferred Reported Item for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines [28].

Search Methodology and Study Selection
The literature search was conducted on Scopus and PubMed databases, and it was limited to English documents. Each database was queried using the following keyword structure: ("wearable" OR "sensors") AND ("ergonomics" OR "occupational medicine" OR "occupational health") AND ("AI" OR "ML").
In order to simplify our research, the exclusion criteria were: • Conference reviews, reviews, book chapters and erratum; • Papers not available; • Papers duplicated.
Concerning the screening by title, abstract, and full text, the following exclusion criteria were defined: • Papers proposing human-machine interface solutions without wearable devices, and not explicitly related to occupational medicine (e.g., touchless control interface in an underwater simulation environment [29]); • Papers proposing wearable devices for cognitive ergonomics (e.g., [30]); • Papers proposing only a wearable device solution without AI (e.g., [31]); • Papers proposing wearable devices for other purposes (e.g., rehabilitation [32]).
Documents were screened evaluating, firstly, title and abstract contents and, in case the documents did not meet the inclusion criteria, secondly the full text. Figure 1 shows the PRISMA workflow, and the number of documents included in this systematic review.

Main Findings and Argumentation
The systematic review includes 25 papers divided into journal articles (16 out of 25), and conference papers (9 out of 25). Most of the papers were published between 2018 and 2021, peaking in 2020 as shown in Figure 2, underlining the growing interest in occupational ergonomics both from a practical and research point of view.

Main findings and Argumentation
The systematic review includes 25 papers divided into journal articles (16 out of 25), and conference papers (9 out of 25).
Most of the papers were published between 2018 and 2021, peaking in 2020 as shown in Figure 2, underlining the growing interest in occupational ergonomics both from a practical and research point of view. The papers were analyzed according to several categories: aim of the study; people involved, and task performed by the subjects; type of wearable device and its positioning on the human body; signal acquired by the sensor and features extracted; principles, methods, standards and/or guidelines underlying the ergonomic assessment; AI strategy (ML and/or DL); results of the studies. Table 1 shows the papers in descending order by year. The papers were analyzed according to several categories: aim of the study; people involved, and task performed by the subjects; type of wearable device and its positioning on the human body; signal acquired by the sensor and features extracted; principles, methods, standards and/or guidelines underlying the ergonomic assessment; AI strategy (ML and/or DL); results of the studies. Table 1 shows the papers in descending order by year.

Wearable Device Type and Study Population
Wearable devices have developed exponentially through novel sensors and technologies, and the long-term monitoring of vital signs and other principles, as described in [58]. The versatility of these devices makes them useful in multiple healthcare scenarios for several purposes (e.g., chronic diseases, mental health and medical conditions) [59][60][61][62][63][64][65]. Authors included studies on wearable solutions for ergonomic risk to prevent WMSDs and suggested two device types: prototype and commercial. Prototype device stands for wearable devices or a system of wearable devices in a configuration not commercially available, while commercial device means commercially available solutions. These included studies recruited healthy subjects, differentiating between volunteers and workers.
Of these studies, 3 out of 25 articles tested a prototype device on healthy volunteer subjects. Akanmu et al. [39] developed an architecture that provides feedback to perform real construction tasks in safe postures. Manjarres et al. [47] suggested a configuration, composed of human activity recognition hardware and a smartwatch, to track physical workload. Low et al. [49] designed a real-time ergonomic risk assessment system to detect workers' movements.
Other prototypes were tested on healthy worker subjects. Specifically, Aiello et al. [35] developed a smart wearable device, placed on wrists, to evaluate vibration risks in industry context, while Campero-Jurado et al. [38] presented a smart helmet to monitor accidents in a work team; finally, Xie and Chang [51] proposed a wearable safety assurance system framework for workers' health in complicated environments. For this last contribution, we showed the system framework in Figure 3. technologies, and the long-term monitoring of vital signs and other principles, as described in [58]. The versatility of these devices makes them useful in multiple healthcare scenarios for several purposes (e.g., chronic diseases, mental health and medical conditions) [59][60][61][62][63][64][65]. Authors included studies on wearable solutions for ergonomic risk to prevent WMSDs and suggested two device types: prototype and commercial. Prototype device stands for wearable devices or a system of wearable devices in a configuration not commercially available, while commercial device means commercially available solutions. These included studies recruited healthy subjects, differentiating between volunteers and workers.
Of these studies, 3 out of 25 articles tested a prototype device on healthy volunteer subjects. Akanmu et al. [39] developed an architecture that provides feedback to perform real construction tasks in safe postures. Manjarres et al. [47] suggested a configuration, composed of human activity recognition hardware and a smartwatch, to track physical workload. Low et al. [49] designed a real-time ergonomic risk assessment system to detect workers' movements.
Other prototypes were tested on healthy worker subjects. Specifically, Aiello et al. [35] developed a smart wearable device, placed on wrists, to evaluate vibration risks in industry context, while Campero-Jurado et al. [38] presented a smart helmet to monitor accidents in a work team; finally, Xie and Chang [51] proposed a wearable safety assurance system framework for workers' health in complicated environments. For this last contribution, we showed the system framework in Figure 3. Some authors tested commercial devices on healthy volunteer subjects. A potential approach is described in [34]; the authors used the Opal (APDM, Inc, USA) [34], which is a wearable inertial system for motion capture composed of several Opal sensors constituted by Inertial Measurement Units (IMUs) [66,67]. Opal sensors communicate through Bluetooth, with a laptop equipped by Mobility Lab Software thanks to the Access Point, while the Docking Station charges and configures sensors. Figure 4 shows the Opal System and the placement of the Opal sensor in the lumbosacral region for the work of Donisi et al. [34]. Some authors tested commercial devices on healthy volunteer subjects. A potential approach is described in [34]; the authors used the Opal (APDM, Inc, USA) [34], which is a wearable inertial system for motion capture composed of several Opal sensors constituted by Inertial Measurement Units (IMUs) [66,67]. Opal sensors communicate through Bluetooth, with a laptop equipped by Mobility Lab Software thanks to the Access Point, while the Docking Station charges and configures sensors. Figure 4 shows the Opal System and the placement of the Opal sensor in the lumbosacral region for the work of Donisi et al. [34].  The Equivital EQ02 Life Monitor system consists of a multi-parameter body worn sensor [40]. Other examples are the Lafayette Hydraulic Hand Dynamometer, a hand dynamometer [44], and the AMS AS7264A, namely a tri-stimulus light color sensor [52].
Finally, Fridolfsson et al. [46] used a commercial shoe-based sensor for classifying work activities on both volunteer, and worker healthy subjects.

Sensor Type and Positioning
The majority of the studies (18 out of 25) used inertial wearable sensors. Inertial sensors refer to accelerometers, gyroscopes and magnetometers that measure linear acceleration, angular velocity and magnetic fields. Typically, three orthogonal gyroscopes, three orthogonal accelerometers, and three orthogonal magnetometers are contained in an IMU [68]. Eight studies combine inertial sensors and other sensor types, as detailed in Table 2.
Finally, Fridolfsson et al. [46] used a commercial shoe-based sensor for classifying work activities on both volunteer, and worker healthy subjects.

Sensor Type and Positioning
The majority of the studies (18 out of 25) used inertial wearable sensors. Inertial sensors refer to accelerometers, gyroscopes and magnetometers that measure linear acceleration, angular velocity and magnetic fields. Typically, three orthogonal gyroscopes, three orthogonal accelerometers, and three orthogonal magnetometers are contained in an IMU [68]. Eight studies combine inertial sensors and other sensor types, as detailed in Table 2. In particular, Matijevich et al. [37] found the best combination of wearable sensors to monitor low back loading during manual material handling, using trunk IMU and pressure insoles.
One of the applications of inertial sensors is the recognition of human postures in several environments, i.e., during activities in the workplace [69]. Posture recognition also depends on where the sensors are attached to anatomical segments of the human body. We divided anatomical segments into three categories to show the body parts mostly considered for the attachment of sensors: "upper body" including the lumbar region, wrist, head, thorax, arm, sternum, pelvis, hip, shoulders, waist and hand; "lower body" including the thigh, shank, calf, foot, leg; and "total body" including both the "upper body" and "lower body". Figure 5 represents inertial sensor distribution according to the three categories proposed. In particular, Matijevich et al. [37] found the best combination of wearable sensors to monitor low back loading during manual material handling, using trunk IMU and pressure insoles.
One of the applications of inertial sensors is the recognition of human postures in several environments, i.e., during activities in the workplace [69]. Posture recognition also depends on where the sensors are attached to anatomical segments of the human body. We divided anatomical segments into three categories to show the body parts mostly considered for the attachment of sensors: "upper body" including the lumbar region, wrist, head, thorax, arm, sternum, pelvis, hip, shoulders, waist and hand; "lower body" including the thigh, shank, calf, foot, leg; and "total body" including both the "upper body" and "lower body". Figure 5 represents inertial sensor distribution according to the three categories proposed. The diagram in Figure 5 shows that most studies placed inertial sensors on the whole body for posture recognition. Three studies [43,46,53] used accelerometers on the lower body only, specifically on feet.
Furthermore, a substantial minority of articles used only biopotential sensors, namely devices that convert a biological response in an electrical signal. In the current study, examples of biopotential wearable sensors were used by Mudiyanselage et al. [33] and Umer et al. [40]. Both authors placed wearable sensors on the upper body, precisely on the thorax, albeit aiming at two different objectives. Specifically, Mudiyanselage et al. [33] studied the level of risk during lifting activities by means of statistical features of an electromyographic signal, as well as Umer et al. [40] that predicted physical exertion levels using statistical features extracted from an electrocardiographic signal. Figure 6 shows the sensors' positioning [40].  The diagram in Figure 5 shows that most studies placed inertial sensors on the whole body for posture recognition. Three studies [43,46,53] used accelerometers on the lower body only, specifically on feet.
Furthermore, a substantial minority of articles used only biopotential sensors, namely devices that convert a biological response in an electrical signal. In the current study, examples of biopotential wearable sensors were used by Mudiyanselage et al. [33] and Umer et al. [40]. Both authors placed wearable sensors on the upper body, precisely on the thorax, albeit aiming at two different objectives. Specifically, Mudiyanselage et al. [33] studied the level of risk during lifting activities by means of statistical features of an electromyographic signal, as well as Umer et al. [40] that predicted physical exertion levels using statistical features extracted from an electrocardiographic signal. Figure 6 shows the sensors' positioning [40]. In particular, Matijevich et al. [37] found the best combination of wearable sensors to monitor low back loading during manual material handling, using trunk IMU and pressure insoles.
One of the applications of inertial sensors is the recognition of human postures in several environments, i.e., during activities in the workplace [69]. Posture recognition also depends on where the sensors are attached to anatomical segments of the human body. We divided anatomical segments into three categories to show the body parts mostly considered for the attachment of sensors: "upper body" including the lumbar region, wrist, head, thorax, arm, sternum, pelvis, hip, shoulders, waist and hand; "lower body" including the thigh, shank, calf, foot, leg; and "total body" including both the "upper body" and "lower body". Figure 5 represents inertial sensor distribution according to the three categories proposed. The diagram in Figure 5 shows that most studies placed inertial sensors on the whole body for posture recognition. Three studies [43,46,53] used accelerometers on the lower body only, specifically on feet.
Furthermore, a substantial minority of articles used only biopotential sensors, namely devices that convert a biological response in an electrical signal. In the current study, examples of biopotential wearable sensors were used by Mudiyanselage et al. [33] and Umer et al. [40]. Both authors placed wearable sensors on the upper body, precisely on the thorax, albeit aiming at two different objectives. Specifically, Mudiyanselage et al. [33] studied the level of risk during lifting activities by means of statistical features of an electromyographic signal, as well as Umer et al. [40] that predicted physical exertion levels using statistical features extracted from an electrocardiographic signal. Figure 6 shows the sensors' positioning [40].  Finally, other studies [44,45,52,56,57] proposed different sensors, such as: skin temperature sensors, respiration sensors, hand dynamometers, pulse oximeters, flex sensors, color light sensors, capacitive sensors, strain sensors, pressure sensors, and inclinometers.

Ergonomic Criteria
Manual material handling is an important risk factor for the development of WMSDs in construction workers. Ergonomic criteria allow the quantification of risk levels during manual handling activities [70], such as those used to design the tasks depicted in Figure 7.
Manual material handling is an important risk factor for the development of WMSDs in construction workers. Ergonomic criteria allow the quantification of risk levels during manual handling activities [70], such as those used to design the tasks depicted in Figure  7.
One of the ergonomic criteria quoted in the systematic literature review is the Revised NIOSH Lifting Equation (RNLE). The RNLE is a manual material handling risk assessment method associated with lifting and lowering tasks in the workplace [71,72]. Mudiyanselage et al. [33] determined three risk classes ("Normal Risk", "Increased Risk" and "High Risk") according to the Revised NIOSH Lifting Equation. All the variables (included in the RNLE) were used to calculate the Recommended Weight Load, and the related Lifting Index (LI) values ranging from 0.8 to 3.2. Similarly, Donisi et al. [34] introduced two risk classes ("No Risk" and "Risk") by combining height, frequency, and weight variables of lifting tasks. They computed two LI values equal to 0.5 and 1.3. Lifting phases of the lifting task are illustrated in Figure 7a, where subjects performed lifting activities using a plastic container with weight equally distributed.  One of the ergonomic criteria quoted in the systematic literature review is the Revised NIOSH Lifting Equation (RNLE). The RNLE is a manual material handling risk assessment method associated with lifting and lowering tasks in the workplace [71,72]. Mudiyanselage et al. [33] determined three risk classes ("Normal Risk", "Increased Risk" and "High Risk") according to the Revised NIOSH Lifting Equation. All the variables (included in the RNLE) were used to calculate the Recommended Weight Load, and the related Lifting Index (LI) values ranging from 0.8 to 3.2. Similarly, Donisi et al. [34] introduced two risk classes ("No Risk" and "Risk") by combining height, frequency, and weight variables of lifting tasks. They computed two LI values equal to 0.5 and 1.3. Lifting phases of the lifting task are illustrated in Figure 7a, where subjects performed lifting activities using a plastic container with weight equally distributed.
Two papers [36,42] classified sensor-detected postures of construction workers, using the Ovako Work Posture Analysis System (OWAS) as a reference. The OWAS method identifies safe/unsafe posture that causes WMSDs [73].
Another ergonomic criterion found in the review, and which was used to prevent ergonomic risk factors, is the Occupational Safety and Health Administration (OSHA) [74].
On one side, Antwi-Afari et al. [43] estimated the ergonomic risk levels ("Low", "Moderate" and "High") according to OSHA by means of the weight of the object, while on the other side Nath et al. [54] estimated the same ergonomic risk levels through the duration and frequency of pushing/pulling, and carrying/lowering/lifting activities. These activities are illustrated in Figure 7b.
Finally, 10 out of 25 articles did not mention a specific ergonomic criterion. For instance, Estrada and Vea [45] classified the sitting posture as ergonomically correct and incorrect, by means of flexible wireless sensors connected to a server. Martire et al. [52] evaluated the ability of AI algorithms to recognize when a user is looking at a digital screen, with a binary classification using features extracted from the sensor. Olsen et al. [57] measured the range of postures of the user, by classifying them as ergonomically correct and incorrect using inclinometers placed on the laboratory coat.

Artificial Intelligence Strategy
ML and DL are two branches of AI that can help to prevent WMSDs, as studied in [17]. In the included articles, the distribution of the methodologies adopted is: 3 studies applied DL, 14 studies applied ML, and 8 studies applied both ML and DL. The most frequently employed algorithms were ensemble classifiers, followed by Support Vector Machines (SVM), Artificial Neural Networks (ANN), k-Nearest Neighbors (kNN), Decision Trees (DTs), generalized linear models, and Naïve Bayes (NB) classifiers. Figure 8 shows the occurrences of the AI algorithms.
Two papers [36,42] classified sensor-detected postures of construction workers, using the Ovako Work Posture Analysis System (OWAS) as a reference. The OWAS method identifies safe/unsafe posture that causes WMSDs [73].
Another ergonomic criterion found in the review, and which was used to prevent ergonomic risk factors, is the Occupational Safety and Health Administration (OSHA) [74]. On one side, Antwi-Afari et al. [43] estimated the ergonomic risk levels ("Low", "Moderate" and "High") according to OSHA by means of the weight of the object, while on the other side Nath et al. [54] estimated the same ergonomic risk levels through the duration and frequency of pushing/pulling, and carrying/lowering/lifting activities. These activities are illustrated in Figure 7b.
Finally, 10 out of 25 articles did not mention a specific ergonomic criterion. For instance, Estrada and Vea [45] classified the sitting posture as ergonomically correct and incorrect, by means of flexible wireless sensors connected to a server. Martire et al. [52] evaluated the ability of AI algorithms to recognize when a user is looking at a digital screen, with a binary classification using features extracted from the sensor. Olsen et al. [57] measured the range of postures of the user, by classifying them as ergonomically correct and incorrect using inclinometers placed on the laboratory coat.

Artificial Intelligence Strategy
ML and DL are two branches of AI that can help to prevent WMSDs, as studied in [17]. In the included articles, the distribution of the methodologies adopted is: 3 studies applied DL, 14 studies applied ML, and 8 studies applied both ML and DL. The most frequently employed algorithms were ensemble classifiers, followed by Support Vector Machines (SVM), Artificial Neural Networks (ANN), k-Nearest Neighbors (kNN), Decision Trees (DTs), generalized linear models, and Naïve Bayes (NB) classifiers. Figure  8 shows the occurrences of the AI algorithms. Ensemble classifiers combine a set of several ML algorithms, named base learners, to obtain a single classifier that outperforms the others [75]. In the included studies, Ensemble classifiers combine a set of several ML algorithms, named base learners, to obtain a single classifier that outperforms the others [75]. In the included studies, ensemble classifiers consist of Random Forest (RF), AdaBoost (AB), Gradient Boost (GB), and Gradient Boost Decision Trees (GBDTs). SVM classifier is a ML technique that creates a gap between the classes, maximizing the distance between them, and reducing misclassification error [76]. ANNs consist of input and output elements called artificial neurons that try to reproduce synaptic links by improving results of conventional algorithms. The output neurons are a weighted sum of input ones [77]. In the present work, ANNs include Multilayer Perceptron (MP), Convolutional Long Short-Term Memory (CLSTM), Static Neural Network (SNN), Convolutional Neural Network (CNN), and Learning Vector Quantization (LVQ). The kNN algorithm is able to make a good classification of an instance if its k-nearest neighbors have the same label. The classification is based on Euclidean distance [78]. DTs represent a sequential structure that divides the data repeatedly, and can be used for the description, generalization and classification of data [79]. The Generalized Linear Models (GLMs) provide a generalization of the linear regression by allowing the linear model to be related to the response variable through a link function [80]. GLMs include Linear and Logistic regression in the systematic review. NB is a probabilistic classifier of supervised learning, based on Bayes' theorem. NB classifiers assume that the value of each feature is independent of the value of any other feature [81].
In terms of accuracy, several classifiers showed high values. Among the ensemble classifiers, RF was the best algorithm showing accuracy values above 90%. Antwi-Afari et al. [43] achieved an accuracy value of 97% in recognizing activities related to overextension, while Manjarres et al. [47] obtained an accuracy value of 97.7% in determining activities performed by volunteer subjects. The best results for ANNs in the classification of awkward positions of workers in terms of accuracy (98.20%) were reached by Antwi-Afari et al. [53]. Another strong result is obtained from Campero-Jurado et al. [38] detecting occupational risks by means of CNN and reaching an accuracy of 92.05 %. With the same number of inertial sensors and an increase of subjects (from 4 to 9), Zhao and Obonyo [36,42] improved the results in terms of Macro F1 score, from 79% to 84%, using CLSTM in the recognition of workers' postures. Accuracy values over 99% were reached by Conforti et al. [41] (99.4%) using the SVM algorithm fed with kinematic data to recognize safe and unsafe postures, and by Olsen et al. [57] (99.94%) to classify ergonomically correct and incorrect postures by means of KNN algorithm.

Feature Extraction
Feature extraction is a useful process of dimensionality reduction and/or redundant data reduction that avoids the loss of important information. The extracted features refer to the signals acquired from sensors placed on a specific body part [82]. These features train ML to classify workers' postures, or to recognize the motor patterns linked to workers' activities.
Two studies [39,41] out of twenty-five used kinematic features extracted from linear acceleration and angular velocity signals as inputs for two different types of techniques, reinforcement learning and supervised learning, respectively. In particular, Conforti et al. [41] used kinematic features (not specified) as SVM inputs that recognized ergonomically correct, and incorrect, with an accuracy value of 99.4%.
Sensor features could be analyzed to identify an ideal wearable system [83]. Matijevich et al. [37] trained several algorithms by means of kinetic and kinematic features in order to find a combination of wearable sensors to monitor low back biomechanical load. The authors used two different sets of wearable sensor signals (idealized wearable sensor signals and real wearable sensor signals) to train ML algorithms. In the ideal configuration, the algorithm identifies the signals that best estimate the lumbar load, i.e., sagittal trunk angle, and vertical ground reaction forces; the real configurations confirm the results of the ideal wearable sensors' signals.
In some articles [34,35,43,46,47,53,54,80,82] statistical features (time and frequency domains features) and spatial-temporal features were extracted from inertial sensor signals. In particular, Donisi et al. [34] extracted time-domain statistical features. After computing feature importance, the authors observed that features associated with acceleration along the y-axis (i.e., mediolateral direction) are more informative to discriminate between two risk conditions, according to the RNLE. The ML results, in particular the RF algorithm, showed an accuracy over 90%. Similarly, Manjarres et al. [47] extracted statistical features (i.e., mean, standard deviation, variance, median absolute deviation) from the linear acceleration signal. The most informative features for RF classifier to track physical workload were the mean of the z-axis (i.e., perpendicular direction to the sensor plane), besides the mean, the standard deviation, and the variance of the x-axis (i.e., vertical direction). In terms of accuracy, RF showed 97.7%.
Differently from the above-mentioned articles, Mudiyanselage et al. [33] extracted statistical features from the surface electromyographic signal. These features trained ML algorithms to classify the risk level of harmful lifting activities with an accuracy of 98%.

Conclusions
The ergonomic analysis technique that makes use of sensors and AI is mainly aimed at the prevention of WMSDs, and particularly affects the body sectors of the upper limbs and back, widely treated in ergonomics. Through this approach, aspects related to the posture of the whole body have also been partly explored, addressed in ergonomics only recently, and for which, in the literature, there are still no clear thresholds of sustainability or indications of optimal levels of variability over time. The application of this approach provides useful information on the needs of ergonomics to improve the conditions of safety at work, and the comfort of the worker; to design suitable work environments and equipment; or to set up work organizations that avoid the onset of phenomena of accumulation of fatigue or overload. Above all, this approach can be advantageous for the analysis of complex or difficult to observe work situations.
As the diffusion of this approach progresses, the wealth of knowledge could help improve the prevention of WMSDs, both associated with acute and cumulative load. This could provide useful information for setting up working methods that are well tolerated, even during the entire working life-an important aspect especially for professions with high biomechanical wear, such as for construction operators or healthcare professionals.
This approach assists not only in the study of the characteristics of force, repetitiveness, and posture (classic risk factors in physical ergonomics), but also in the kinematic traits of the worker's behavior. Specific kinematic traits could be useful as indicators to control and predict the appearance of any alterations capable of endangering the integrity of the worker, but also to monitor the critical phases during the return to work for people with dysfunctions, disabilities or previous pathologies.
Furthermore, the data detectable through sensors can enrich the value of the ergonomic intervention of evaluation and design, attracting interest also on aspects properly investigated by other disciplines, such as engineering, psychological, organizational, medical, but also economic ones. The technological approach can be all the more innovative the more it uses prototypes (rather than commercial standard tools), often made with open-source resources, and not pre-deterministically channeled towards a single aspect of interest. Considering some variables detectable through sensors, the design of optimal work situations can be addressed to specific categories of workers, such as the elderly, in order to be able to implement targeted adaptations of the workplace that guarantee the expected levels of productivity and safety.
In addition to the purposes of monitoring, evaluation, and design, the combined technique that uses sensors and AI opens up new scenarios for ergonomic interventions of an educational and participatory prevention type; this provides a contribution for workers to explore new ways of carrying out work, possibly also with the adoption of technological aids and devices, such as exoskeletons. The illustrated approach also opens the way to analysis and consideration of multiple conditions of exposure to physical, chemical, environmental, organizational factors at work, for which neither consolidated methodologies for risk assessment are currently available nor is evidence of association available, with the motor, physiological or biomechanical functions of the human operator.
Further studies may make improvements to the illustrated technique, specifying the optimal positioning of the sensors, defining the best AI system, but also proposing the elaboration and development of other methods of ergonomic analysis, different from those already used and accepted by classical ergonomics. An interesting aspect of the study related to the topic presented here, and mainly focused on WMSDs, concerns the interpretation of worker well-being as an integrated construct that includes physical, psychosocial, and organizational aspects (1948 WHO definition of health). As it has, in fact, been demonstrated by various studies, these aspects act with reciprocal influence on the conditions of the human operator, and the intervention on one of the risk factors could have repercussions on the other dimensions. This broadening of perspective also affects the long-term benefits that can be prepared for, and guaranteed by, short-term investments in improving occupational safety and health. Furthermore, given the multifactorial nature of the underlying causes of WMSDs, a future study perspective could concern the assessment of exposure associated with prolonged low-intensity static work, typical of teleworkers and the increasing digitalization of work.
This article presented a systematic review of the combined use of wearable devices and AI for ergonomic purposes, selecting 25 relevant studies from the scientific literature. The analysis highlighted a deep interest, which has grown in recent years, for the use of wearable sensors coupled with AI algorithms (both ML and DL) to monitor the biomechanical risk to which workers are exposed to during their activities. The review provides the researcher with an overview of the latest uses of AI and wearable sensors in the context of physical ergonomics. Additionally, this review could be useful to support professionals in selecting the most suitable wearable technology and AI strategy for ergonomic assessments and improvements in industrial and non-industrial settings.