Accelerometer-Based Identification of Fatigue in the Lower Limbs during Cyclical Physical Exercise: A Systematic Review

Physical exercise (PE) is beneficial for both physical and psychological health aspects. However, excessive training can lead to physical fatigue and an increased risk of lower limb injuries. In order to tailor training loads and durations to the needs and capacities of an individual, physical fatigue must be estimated. Different measurement devices and techniques (i.e., ergospirometers, electromyography, and motion capture systems) can be used to identify physical fatigue. The field of biomechanics has succeeded in capturing changes in human movement with optical systems, as well as with accelerometers or inertial measurement units (IMUs), the latter being more user-friendly and adaptable to real-world scenarios due to its wearable nature. There is, however, still a lack of consensus regarding the possibility of using biomechanical parameters measured with accelerometers to identify physical fatigue states in PE. Nowadays, the field of biomechanics is beginning to open towards the possibility of identifying fatigue state using machine learning algorithms. Here, we selected and summarized accelerometer-based articles that either (a) performed analyses of biomechanical parameters that change due to fatigue in the lower limbs or (b) performed fatigue identification based on features including biomechanical parameters. We performed a systematic literature search and analysed 39 articles on running, jumping, walking, stair climbing, and other gym exercises. Peak tibial and sacral acceleration were the most common measured variables and were found to significantly increase with fatigue (respectively, in 6/13 running articles and 2/4 jumping articles). Fatigue classification was performed with an accuracy between 78% and 96% and Pearson’s correlation with an RPE (rate of perceived exertion) between r = 0.79 and r = 0.95. We recommend future effort toward the standardization of fatigue protocols and methods across articles in order to generalize fatigue identification results and increase the use of accelerometers to quantify physical fatigue in PE.


Introduction
Physical exercise (PE) benefits human beings in many ways: from a psychological perspective, reducing anxiety and risk of depression [1]; from a physiological perspective, avoiding a sedentary lifestyle and reducing risks of cardiovascular diseases [2].; from a biomechanical perspective, reducing risk of musculoskeletal disorders (MSDs) [3]; and from a neurological perspective, improving cognitive functioning and counteracting aging processes [1]. However, PE can also lead to injuries, especially when PE activities are Sensors 2022, 22, 3008 3 of 36 In this study, we aim to contribute to state-of-the-art PE monitoring with a comprehensive overview of the performance of accelerometer-based methods to identify fatigue in cyclical PE, since cyclical tasks allow for a comparison across different PE activities. A literature search is performed including articles that assess biomechanical changes in fatiguing cyclical PE activities or use such changes to identify a fatigued state. We aim to provide an overview of the literature regarding accelerometer-based measures of biomechanical changes due to fatigue, as well as an overview of the literature regarding the detection of fatigue via models or machine learning approaches that use kinematic features measured via accelerometers.

Search Strategy
This review was reported following the PRISMA guidelines (Table S1, Supplementary Materials) [34]. An exhaustive search strategy was developed by an experienced information specialist (WMB). The search was developed in Embase.com, optimized for sensitivity, and then translated to other databases following the method described by Bramer et al. [35]. The search was carried out in the following databases: Embase.com (date of inception 1971), Medline ALL via Ovid (1946 to Daily Update), Web of Science Core Collection, and CINAHL Plus via EBSCOhost. After the original search was performed on 5 August 2020, the search was last updated on 31 May 2021 using the methods described by Bramer et al. [36].
The search strategies for Embase and Medline used relevant thesaurus terms from Emtree and Medical Subject Headings (MeSH), respectively. In all databases, terms were searched in the titles and abstracts of references. The search contained terms for: (1) fatigue or exhaustion; (2) physical exercise, gait, walking, or running; and (3) inertial measurements or accelerometry. Terms were combined with the Boolean operators AND and OR, and proximity operators were used to combine terms into phrases. The full list of the keywords used in each search strategy for all four databases is available in Appendix A (Table A1). The searches in Embase and Web of Science were limited to exclude conference papers older than 3 years. In all databases, non-English articles and animal-only articles were excluded from the search results. No study registries were searched. The reference lists of retrieved nonincluded relevant review articles and of the included references, as well as articles citing these papers, were scanned for relevant references missed by the search. No authors or subject experts were contacted, and we did not browse unindexed journals in the field. The references were imported into EndNote, and duplicates were removed by the medical librarian (WMB) using the method described by Bramer et al. [37].

Screening of Articles and Eligibility Criteria
Two reviewers (LM and BS) independently screened titles and abstracts in EndNote using the method described by Bramer et al. [38]. Any discrepancies in the verdict were resolved by discussion with a third reviewer (JR). A total of seven exclusion criteria in the abstract screening phase (Table A2) and eight exclusion criteria in the full-text screening phase (Table A3) were used and can be found in Appendix A. Two reviewers (LM and BS for the first half of the articles in alphabetical order, and LM and RvM for the second half of articles) independently screened the full-text articles. Any discrepancies in the verdict were resolved by discussion with a third reviewer (JR).
The aim of this review was to assess accelerometer-based methods to identify fatigue in cyclical PE. In the initial search strategy, work activities were still included. However, the recovery time and intensity of such activities have very large variations compared to cyclical individual PE activities (i.e., running, walking, jumping, and stair climbing). Furthermore, work task movement patterns lack continuity when compared to cyclical PE tasks and were, therefore, excluded (EC2.2, Table A3).

Data Extraction
A total of 2889 articles were retrieved, resulting in a total number of 39 articles included in this review after the screening process ( Figure 1). After the initial search (5 August 2020), removal of duplicates, screening of titles, and abstract and screening of full-texts, thirty articles were included (Supplementary Material, Figure S1). After performing the search a second time (31 May 2021), eight new articles were identified (Supplementary Material, Figure S1). One article was identified through citation searching.

Data Extraction
A total of 2889 articles were retrieved, resulting in a total number of 39 articles included in this review after the screening process ( Figure 1). After the initial search (5 August 2020), removal of duplicates, screening of titles, and abstract and screening of fulltexts, thirty articles were included (Supplementary material, Figure S1). After performing the search a second time (31 May 2021), eight new articles were identified (Supplementary Material, Figure S1). One article was identified through citation searching.  Table A2. ** Records excluded via exclusion criteria in Table A3.  Table A2. ** Records excluded via exclusion criteria in Table A3.

Outcomes of Interest
All articles included in this review aimed to identify fatigue in the lower limbs during PE using accelerometers. Articles that aimed to quantitatively identify changes due to fatigue in lower limb biomechanics were classified as Type I. The outcomes of interest for these articles were biomechanical parameters measured before and after a fatiguing intervention. Biomechanical parameters were kinematic or spatiotemporal variables that can be measured directly using accelerometers: segment accelerations, shock attenuation, and stride parameters. Articles that aimed to identify, predict, or classify fatigue states based on quantitative biomechanical features were classified as Type II. The outcomes of interest for Type II articles were the performance metrics of the proposed model or classifier. Additional variables for which data were sought concerned study protocol, subject population, measurement system, intervention (fatigue protocol), fatigue reference, and data analysis techniques.

Segment Accelerations
The lower limb accelerations considered in this review were feet, tibia, thigh, and sacrum segmental accelerations. Peak segmental accelerations are commonly used to understand human motion, and they have been linked, in particular, to tibial bone loading [39], which could provide relevant information in understanding injury risk. Peak tibial accelerations are commonly used as an indirect measure of impact during running [27].

Shock Attenuation
Shock attenuation is the magnitude or frequency reduction of the impact shock wave in human movements that involve an impact of the lower limbs with the ground [40]. Shock attenuation strategies are used by the body to deal with high impacts with the ground that can happen during various PE activities [27]. The shock attenuation outcomes considered in this review were between the tibia and sacrum, trunk, or forehead.

Stride Parameters
Stride spatiotemporal parameters are commonly used to describe the human gait (e.g., stride frequency, stride length, and stride time). Stride parameters can be related to cumulative load and contain relevant information to prevent running-related injuries [41], and were, therefore, considered in this review.

Model Performance Metrics
For Type II articles, models were built to identify or classify fatigue states. The performances of these models were evaluated by means of accuracy metrics, typically used in classification problems, or correlation metrics, typically used in regression problems. The accuracy, sensitivity, and specificity of the classifier are common performance metrics in classification problems, while root mean squared error (RMSE) and Pearson's r are frequently used correlation metrics.

Quality Assessment
A quality assessment checklist was adapted from the Downs and Black checklist [42], tailoring criteria regarding reporting, internal and external validity, and power. Twelve items were selected and used for Type I articles. Additionally, items adapted from the Luo et al. Guidelines for Developing and Reporting Machine Learning Predictive Models in Biomedical Research [43] were added for the quality assessment of Type II articles. Four items were selected and used to replace Downs and Black items in order to tailor the checklist to Type II articles. The full quality assessments for Type I (Table S2) and Type II (Table S3) articles are reported in the Supplementary Materials. Selected items from the two quality assessment checklists are shown in Table 1. The maximum score for both types of articles was 11; articles that did not exceed a threshold of 5/11 were discarded.

Overview of Article Characteristics
Articles that passed the screening phase and were included in the review are shown in Table 2. The article identification process led to 39 articles. PE activities included running (28 articles), walking (4 articles), jumping exercises (4 articles), stair-climbing tests (SCTs) (2 articles), and gym exercises (1 article). The aim of this section is to summarize the subject populations, sensor placements, fatiguing protocols, fatigue references, and outcomes of interest for the included articles.

Measurement System and Sensor Placement
The measurement systems used to identify biomechanical parameters were simple accelerometers (20 articles) or accelerometers embedded in IMUs (19 articles). In three cases, the accelerometers or IMUs were embedded in a smartphone.
The accelerometers were fixed to a single body segment in 16 articles (10 tibia, 5 sacrum only, and 1 foot) and to multiple body segments in 23 articles. In particular, the tibia was chosen as a sensor location in 78% of the running articles and 79% of all articles, while the percentage of the sacrum placement was consistent at 46%. The foot and thigh were chosen as sensor locations for only 26% and 13%, respectively, out of all articles. A summary of the sensor placement for all articles and each activity can be found in Table 3. Table 3. Accelerometer placement: absolute number and percentage.

Running
Jumping Walking SCT Gym Exercises Total The accelerometers were attached on both limbs in only six articles. Placement of the accelerometer was reported also for the forehead (10 articles) and sternum (6 articles), since they are needed for computing shock attenuation. Measurement systems characteristics and placement are reported in detail in Table 2.

Fatiguing Protocol
Fatiguing protocols varied across the articles. A comprehensive summary of the measured activities and related fatiguing protocols across all the articles can be found in Figure 2. A total of 30/39 articles (77%) reported a fatiguing protocol consisting of the same activity as the main measured activity. All six articles that reported walking and SCT as the main measured activity used a different activity as a fatiguing protocol.
Sensors 2022, 22, x FOR PEER REVIEW Figure 2. Summary of fatiguing protocols and measured activities. Circles on the left side the number of articles that measured each PE activity, while circles on the right side rep number of articles for each PE activity chosen as a fatiguing protocol. Horizontal arrows the articles that used the same fatiguing protocol PE activity as the measured activ diagonal arrows represent the articles that chose a different PE activity.  Table 4.        The stopping criteria of the fatiguing protocols varied per PE activity. Out of 30 articles that reported a running fatiguing protocol, nine articles based the stopping criteria on the length of the run; nine articles based it on a threshold for RPE, HR, or end-tidal carbon dioxide pressure (PETCO2); five articles let participants run until exhaustion; and three articles based their stopping criteria on a decrease in performance, while four articles did not report clear stopping criteria. A decrease in performance was also used as stopping criterium for the two articles reporting a jumping fatiguing protocol and three articles reporting squatting, recumbent ergometer, and triceps surae fatiguing protocols. The only protocol based on gym exercises used a stopping criterion of subject exhaustion, while the remaining two articles reporting squatting fatiguing protocols used an RPE threshold as the stopping criterion. Finally, one recumbent ergometer fatiguing protocol was reported to be stopped when subjects felt uncomfortable.

Fatigue Reference
The fatigue reference metrics across all the articles are reported in Table 5. A total of 19 out of 39 articles (49%) reported RPE as a fatigue reference. The RPEs consisted solely of Borg's RPE [9] (either on a 6-20 or 1-10 scale) for all fatiguing protocols, except for the recumbent ergometer, which also included perceived muscle soreness as a fatigue reference. The HR parameters consisted of changes in HR, absolute HR values, and relative changes compared to the HR max and accounted for 15% of all articles. The ventilatory parameters consisted of changes in PETCO2 and VO2 max, accounting for 18% of articles. Other physiological parameters included changes in creatine kinase and blood lactate concentration. A total of 5 out of 39 articles combined multiple fatigue references in their protocols.

Outcomes of Interest
A total of 32 articles evaluated changes due to fatigue in lower limb biomechanics (Type I), and 7 articles used machine learning approaches to identify, classify, or predict fatigue stages (Type II). Performance metrics were chosen by all seven Type II articles and are presented in Section 3.4.
For Type I articles, peak tibial acceleration (PTA) was the most common reported outcome, chosen in 13 running articles and 3 jumping articles. Shock attenuation was reported in 11 articles: seven times between head and tibia (six running, one jumping) and two times between sacrum and tibia (running). Peak sacral acceleration (PSA) was reported in five articles (four running, one jumping) and peak foot acceleration (PFA) in one running article. Other acceleration-based variables that were reported in running articles were tibial acceleration reduction, tibial impact rate, and peak-to-peak tibial acceleration.
Stride spatiotemporal parameters were chosen both in running and jumping articles. Stride length was the most common variable (six articles), followed by step frequency, stride frequency, stride time, and contact time (each reported in two articles).
Step length and foot strike angle (FSA) were also reported in one article each.
Other variables were also chosen in different activities. In running, some articles focused on frequency domain parameters (i.e., local dynamic stability, power spectral density, and signal power magnitude). Center of mass (COM) displacement was chosen in one running article, while one jumping article reported vertical displacement of the sacrum. Other reported outcomes in jumping articles were touchdown angle, peak tibial angular velocity, and maximal vertical velocity and acceleration of the sacrum. In SCTs, the ranges of motion of the ankle, knee, thigh, and trunk were reported.

Quality Assessment
Quality assessment scores for each article are shown in Table 4. All 39 articles that were evaluated after full-text screening exceeded the threshold of 5/11. The overall quality assessment score was 9.3 ± 1.3 (9.2 ± 1.3 for Type I and 9.6 ± 1.1 for Type II articles). A complete assessment of all the quality assessment items for each article can be found in the Supplementary Materials (Tables S2 and S3).

Running
The changes due to fatigue in biomechanical parameters for running activities can be found in Table 6. Increasing PTA with fatigue was found in 11/13 articles. In 2/13 articles, PTA increased or decreased with fatigue depending on different conditions (i.e., running environment and shoe characteristics). A total of 6/13 articles found significant increases of PTA with fatigue, while 1/13 articles found a significant decrease of PTA. PSA was found to increase with fatigue in 4/4 articles, although only 2/4 articles found the increase to be significant. PFA was found to increase with fatigue by 1/1 article, although the increase was significant only for recreational runners. Shock attenuation between the head and tibia increased in 3/6 articles, while in 2/6 articles, it was found to increase depending on different conditions (i.e., shoe characteristics) or mathematical calculations (transfer function vs. ratio). A total of 2/6 articles found a significant increase in head-to-tibia shock attenuation with fatigue, while 1/6 articles found a significant decrease. Furthermore, 2/2 articles found an increase in sacrum-to-tibia shock attenuation with fatigue, one of them being significant. Significant changes in stride and step spatiotemporal parameters were found in 5/16 articles (1/6 found significant increase in stride length; 1/2 in stride frequency; 0/2 in stride time; 1/2 in step frequency; 1/2 in contact time; 0/1 in step length; 1/1 found significant decrease in FSA).
Significant changes in fatigue reference between the fatigued and non-fatigued states were found in all five articles that reported them. A significant increase in RPE with fatigue was found in one article; a significant increase in oxygen consumption was found in one article; a significant increase in heart rate was found in one article; and a significant decrease of end-tidal carbon dioxide pressure was found in two articles. The average RPE in the fatigued state was reported by four articles and was equal to 15.7 ± 1.4 (6-20 Borg Scale).

Walking
All four walking articles were categorized as Type II articles and are reported in Section 3.4.

Stair-Climbing Test
A total of 1/2 SCT articles investigated changes in biomechanical parameters. An increase in ankle, knee, thigh, and trunk range of motion (ROM) was found with fatigue during descent, the trunk ROM being the only one showing a significant difference. Nonsignificant increases were found with fatigue in knee, thigh, and trunk ROM during ascent, while a non-significant decrease with fatigue was found in ankle ROM.

Jumping Exercises
Changes due to fatigue in biomechanical parameters for jumping activities can be found in Table 7. A total of 3/3 jumping articles found an increase in PTA with fatigue. A significant increase in PTA with fatigue was found in one article only in shorter jumps (30 cm), while it was non-significant in higher jumps (50 cm). Another article found a significant increase in PTA with fatigue in landing, but a non-significant increase during take-off. A total of 1/1 article found a significant increase in PSA with fatigue. A total of 0/1 articles found a significant increase or decrease in head-to-tibia shock attenuation with fatigue in jumping.

Overview of Fatigue Classification Performances
The model characteristics and classification performance for Type II articles can be found in Table 8. A total of 3/4 articles that investigated fatigue in walking used machine learning models, obtaining an accuracy ranging between 78% and 96%. A support vector machine (SVM) model was chosen in all three articles, in one case being the best-performing model compared to multiple different machine learning models. A total of 1/4 articles used multivariate forecast models to predict fatigue states. The best-performing model was an autoregressive integrated moving average (ARIMA) model with a mean absolute error (MAE) of 0.73 with respect to measured RPE values.
In SCTs, 1/1 article developed a model based on changes in body postures and kinetic energy to output a fatigue score. Correlation of the fatigue score with the RPE was quantified by Pearson's r, being equal to 0.95 for males and equal to 0.70 for females.
In gym exercises, 1/1 article used machine learning models to estimate RPE values. Correlation between model outputs and the RPE was quantified by means of Pearson's r, showing different results for different gym exercises: r = 0.89 for squats, r = 0.93 for jumping jacks, and r = 0.94 for corkscrew exercises. The machine learning models used were convolutional neural networks (CNN) and random forest (RF).   In running, 1/1 article developed a multiple linear regression time-to-exhaustion model. Pearson's r was used to quantify correlation between the model's output and RPE, with r being equal to 0.792.
Feature importance analyses were performed in 4/7 Type II articles across all the activities. In two articles, feature performance was performed before training the final model in order to improve model performance. In two articles, feature importance of the model was shown for the final model.

Discussion
The main scope of this literature review was to assess whether accelerometers are suitable sensors to identify physical fatigue in PE. In order to understand the real-life possibilities of fatigue detection in PE, we aimed to assess the capability of accelerometerbased parameters to straightforwardly estimate (traditional biomechanics) and assist in the detecting (machine learning) of physical fatigue. We found that identification of fatigue in PE using inertial sensors is mainly obtained by a straightforward comparison of biomechanical variables of interest or by training models that are validated by comparisons with physiological or subjective fatigue references.
Peak tibial and sacral acceleration were the most commonly sought outcomes. An increase in peak tibial or sacral acceleration with fatigue was found in 19/21 articles for running and jumping activities. However, segment acceleration was influenced by subject characteristics and the type of fatigue protocol (at particular speeds). Reporting these characteristics would facilitate the normalization of segmental acceleration results across articles and provide general, rather than individual, insight in its changes due to fatigue in PE. Other factors that were found to influence segment accelerations are training experience (elite vs. recreational), shoe type (prefabricated vs. custom-made sole), and running environment (treadmill vs. overground). This could explain the high variability across articles on PTA (4.5-24.6 g). Shock attenuation was found to increase with fatigue in 5/9 articles (running and jumping). While a high variability in biomechanical variables due to subject characteristics, number of subjects, and fatiguing protocols did not allow general conclusions, accelerometers were able to measure peak accelerations and shock attenuations at an individual level. Stride spatiotemporal parameters were also measured by accelerometers at an individual level in running, and significant changes were found in 5/16 articles. The low amount of articles that found a significant change in spatiotemporal parameters can be explained by the controlled constant speed in the majority of them.
Identification of physical fatigue using machine learning or other types of algorithms was performed in only 7 out of 39 articles. The accuracy of the models ranged between 78% and 96%, and Pearson's correlation with RPE ranged between 0.79 and 0.95. Only two articles performed cross-validation (k-fold), suggesting that the validity of their results was specific for their subject population. Four articles provided further interpretation of their results by means of feature importance analysis, although the choice of features was either subjective or not specified. Changes in biomechanical variables found in the literature could provide a more objective choice of features for machine learning classifiers. While a generalized optimal method for PE activities was not found in this review, machine learning approaches succeeded in lower limb fatigue identification for each specific activity and were found to be less influenced by fatigue protocol characteristics than traditional biomechanics approaches.
Currently, a gold standard for the comprehensive measurement of physical fatigue in PE is missing. A total of 19/39 articles used Borg's RPE as a fatigue reference or tried to predict and detect RPE levels. Borg's RPE is a very practical scale to estimate fatigue, but it relates only to the mental components of fatigue. A total of 13/39 articles used cardiovascular or ventilatory parameters as fatigue references. They have the advantage of being objective metrics, but they are individual, often difficult to measure outside of a lab, and mostly related to the cardiovascular components of fatigue. Accelerometers have the potential to become extremely popular devices in the identification of physical fatigue in prolonged tasks out of a lab, but research protocols in the field of fatigue identification in human movement and PE are still too different from each other to draw general conclusions. Therefore, we provide five recommendations for future research in PE that could also be generally applied to human movement assessment (e.g., team sports, rehabilitation, and clinical practice) and may help the validation of accelerometers as a measurement system for the identification of physical fatigue.

1.
One of the aims of this review was to assess to what extent the fields of biomechanics and machine learning are useful to each other in fatigue identification. While a few articles developed fatigue models and assessed changes in biomechanics or feature importance [57,68,74], there is still uncertainty in the choice of model and machine learning biomechanical features. Developing consistent fatiguing protocols and reporting feature performance would improve biomechanical domain knowledge in machine learning studies, while automatic feature extraction techniques could also be used to improve model performance, as advocated by Halilaj et al. [29].

2.
Biomechanical parameters of interest for fatigue estimation are influenced by many variables in PE. In this review, we identified sensor location, fatigue protocol, subject characteristics, training level, equipment, and environment. For example, accelerometer location on the distal part of the tibia causes an exposure to higher impacts and higher PTA than the proximal tibia. Accurate method descriptions would allow the proper comparison of biomechanical parameters and the generalization of results.

3.
A subject being either in a fatigued or non-fatigued state is a simplified representation of more complex fatigue models that occur at the cardiovascular and neuromuscular levels [75]. An effort should be made in understanding and identifying fatigue development stages throughout a PE activity.

4.
Fatigue detection, identification, or prediction with machine learning techniques should be generalized over subjects unless the objective is to train a subject-specific model [29]. Fatigue identification in PE is a large-scale problem and should be tackled with a subject-general model, since subject-specific models have limited scalability [76].
Leave-one-subject-out cross-validation should be used when trying to detect outcomes from different subjects, since it significantly helps model performance on new, unseen subjects [28].

5.
Deep learning algorithms were not found in this review, although deep learning could be a promising technique to improve fatigue identification performance by reducing the need for feature engineering [29]. A possible explanation for the lack of deep learning algorithms could be the limited amount of data to train a deep leaning model with a good performance. The online sharing of data across research articles (also advocated by Gurchiek et al. [76]) could help developing a large dataset of accelerometer-based fatigue measurements in each PE activity.

Limitations
The main limitation of this study was the bias towards running activities (thirty out of thirty-nine articles). A possible explanation is the widespread popularity of running as a PE activity and the fact that its cyclical nature makes it an easy activity to analyze in research. However, the biomechanical outcomes of running can be applicable to other PE activities due to their quasi-cyclical nature [77], as well as more complex activities such as team sports. Team sports were not in the scope of this review, but running (and jumping, also evaluated in this review) are predominant component in many of them. Furthermore, extensive research in measuring running biomechanics using IMUs could be used by other sports that are starting to use a similar approach to monitor athletes in order to not repeat the same mistakes.
A second limitation in the analysis of accelerometer-based techniques was the assumption of similarity between accelerometer and IMU measurements. Although IMUs integrate data from gyroscopes and magnetometers, we assumed a neglectable impact on the measured outcomes of interest in our review. Further research would be needed to fully understand whether measurements performed with IMUs differ from measurements performed with simple accelerometers.
A third limitation was the lack of uniformity in fatigue protocols between the articles of this review. Fatigue protocols with different intensities (e.g., higher vs. lower speed) or different activities (e.g., running vs. squatting) can impact muscle activation differently. Single-muscle fatigue assessment was not in the scope of this review, but it has an impact on the onset of overall physical fatigue. Future studies should investigate the possibility to identify physical fatigue levels and link them to activity intensity. A standardization of fatigue protocols could also allow a meta-analysis of changes in biomechanical variables with fatigue in PE.

Conclusions
We aimed to assess whether accelerometer-based techniques could identify lower limb physical fatigue in PE. We found that changes in biomechanical parameters could be assessed at an individual level due to fatigue and that machine learning could help detect fatigue, but the link between machine learning and changes in biomechanics needs to be further investigated. Therefore, we formulated guidelines for future fatigue identification research using accelerometers. The aligning of fatigue protocols and online sharing of data could help validate biomechanical changes due to fatigue in the lower limbs and the large-scale deployment of accelerometers in physical fatigue assessment during PE.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/s22083008/s1: Table S1: PRISMA 2020 Checklist; Figure S1: Flow of systematic review process according to PRISMA diagram after first search (August 2020); Figure S2: Flow of systematic review process according to PRISMA diagram after second search (May 2021); Table S2: Full Quality assessment for Type I articles; Table S3: Full Quality assessment for Type II articles.  limb* OR extermit*)) OR knee OR knees OR hip OR hips OR ankle* OR foot OR feet OR leg OR legs OR thigh* OR work-task* OR (daily N2 (life OR living) N5 activit*) OR ((physical* OR Motor*) N2 activit*) OR stride OR kinetic* OR motion OR biomechanic* OR treadmill* OR exercise* OR work OR workplace OR worker* OR stand OR standing) OR ab(gait OR walking OR running OR jogging OR (lower N2 (limb* OR extermit*)) OR knee OR knees OR hip OR hips OR ankle* OR foot OR feet OR leg OR legs OR thigh* OR work-task* OR (daily N2 (life OR living) N5 activit*) OR ((physical* OR Motor*) N2 activit*) OR stride OR kinetic* OR motion OR biomechanic* OR treadmill* OR exercise* OR work OR workplace OR worker* OR stand OR standing)) AND (MH Accelerometry OR (MH "Acceleration (Mechanics)" AND (MH Smartphone OR MH Mobile Applications)) OR TI(((inertial*) N2 measur*) OR acceleromet* OR gyroscope* OR imu OR imus OR immu OR immus OR ((inertial OR body OR wearable*) N1 sens*) OR xsens OR x-sens OR jerk OR ((smartphone* OR app OR mobile-application*) N2 (accelerat* OR Measure*))) OR AB(((inertial*) N2 measur*) OR acceleromet* OR gyroscope* OR imu OR imus OR immu OR immus OR ((inertial OR body OR wearable*) N1 sens*) OR xsens OR x-sens OR jerk OR ((smartphone* OR app OR mobile-application*) N2 (accelerat* OR Measure*)))) AND LA(English) NOT (MH animals+ NOT MH humans+) Web of science Core Collection (Last accessed on 31 May 2021) TS = (((fatigue* OR exhaust* OR exertion* OR tired*)) AND ((gait OR walking OR running OR jogging OR (lower NEAR/2 (limb* OR extermit*)) OR knee OR knees OR hip OR hips OR ankle* OR foot OR feet OR leg OR legs OR thigh* OR work-task* OR (daily NEAR/2 (life OR living) NEAR/5 activit*) OR ((physical* OR Motor*) NEAR/2 activit*) OR stride OR kinetic* OR motion OR biomechanic* OR treadmill* OR exercise* OR work OR workplace OR worker* OR stand OR standing)) AND ((((inertial*) NEAR/2 measur*) OR acceleromet* OR gyroscope* OR imu OR imus OR immu OR immus OR ((inertial OR body OR wearable*) NEAR/1 sens*) OR xsens OR x-sens OR jerk OR ((smartphone* OR app OR mobile-application*) NEAR/2 (accelerat* OR Measure*))))) AND DT = (article) AND LA = (english) Table A2. Eligibility criteria (EC) during title and abstract screening phase. Articles were excluded if the title or abstract suggests that: The study population is formed by non-healthy subjects, either at the time of the study or in rehabilitation EC 1.2 The study population average age is lower than 18 years or higher than 70 years EC 1.3 The activities performed are not a sport, ADL, or working task involving moving or standing EC 1.4 The study does not include one of the three following sensors: accelerometer, gyroscope, or magnetometer, or does not include IMUs EC 1.5 The study does not include at least a sensor on the lower limbs EC 1. 6 The article is not in English EC 1.7 The article is a conference paper The study does not focus on individual physical exercise tasks EC 2. 3 The study performed measurements spreading over multiple days EC 2.4 The study protocol requires subject to perform power-assisted body movements EC 2.5 The study does not include kinetic or kinematic parameters for the lower limbs EC 2.6 The study lacks a fatigue inducement protocol, in particular: For protocols involving running: no running-induced fatigue (minimum of no-stop 3 km running if not stated) For protocols involving static physical exercise or walking: protocol-induced fatigue and exhaustion EC 2.7 The study protocol includes an accelerometer or IMU with a sampling frequency <60 Hz EC 2.8 The study is a case study (1 subject only)