Towards Machine Learning-Based Detection of Running-Induced Fatigue in Real-World Scenarios: Evaluation of IMU Sensor Configurations to Reduce Intrusiveness

hysical fatigue is a recurrent problem in running that negatively affects performance and leads to an increased risk of being injured. Identi cation and management of fatigue helps reducing such negative effects, but is presently commonly based on subjective fatigue measurements. Inertial sensors can record movement data continuously, allowing recording for long durations and extensive amounts of data. Here we aimed to assess if inertial measurement units (IMUs) can be used to distinguish between fatigue levels during an outdoor run with a machine learning classi cation algorithm trained on IMU-derived biomechanical features, and what is the optimal con guration to do so. Eight runners ran 13 laps of 400 m on an athletic track at a constant speed with 8 IMUs attached to their body (feet, tibias, thighs, pelvis, and sternum). Three segments were extracted from the run: laps 2–4 (no fatigue condition, Rating of Perceived Exertion (RPE) = 6.0 0.0); laps 8–10 (mild fatigue condition, RPE = 11.7 2.0); laps 11–13 (heavy fatigue condition, RPE = 14.2 3.0), run directly after

athletic track at a constant speed with 8 IMUs attached to their body (feet, tibias, thighs, pelvis, and sternum). Three segments were extracted from the run: laps 2–4 (no fatigue condition, Rating of Perceived Exertion (RPE) = 6.0 0.0); laps 8–10 (mild fatigue condition, RPE = 11.7 2.0); laps 11–13 (heavy fatigue condition, RPE = 14.2 3.0), run directly after a fatiguing protocol (progressive increase of speed until RPE 16) that followed lap 10. A random forest classi cation algorithm was trained with selected features from the 400 m moving average of the IMU-derived accelerations, angular velocities, and joint angles. A leave-one-subject-out cross validation was performed to assess the optimal combination of IMU locations to detect fatigue and selected sensor con gurations were considered. The left tibia was the most recurrent sensor location, resulting in accuracies ranging between 0.761 (single left tibia location) and 0.905 (all IMU locations). These ndings contribute toward a balanced choice between higher accuracy and lower intrusiveness in the development of IMU-based fatigue detection devices in running. Keywords:fatigue estimation; biomechanics; IMU; machine learning; human movement; running 1. Introduction Running is an increasingly popular sport, with multiple health bene ts. Fifty million Europeans engage in running according to a recent estimate [1]. Health bene ts can be psychological, such as a sense of accomplishment [2], or physical, such as a decreased chance of developing chronic diseases and higher longevity [3]. However, running comes with associated risks, in particular pain and injuries [4]. To maximize health bene ts and minimize chances of being injured, load should be carefully managed. For instance, overloading and training stress are associated with increased injury risk [5,6]. Accurate, continuous detection of fatigue during a high intensity or long duration running activity can be used to provide feedback to runners in order to avoid excessive training stress and overloading which can lead to lower limb injuries. Fatigue is a multi-factorial phenomenon. A model developed by Kluger et al. [7] divides fatigue in two distinct components that have the capacity to in uence each other: Sensors2021,21, 3451.

used to provide feedback to runners in order to avoid excessive training stress and overloading which can lead to lower limb injuries. Fatigue is a multi-factorial phenomenon. A model developed by Kluger et al. [7] divides fatigue in two distinct components that have the capacity to in uence each other: Sensors2021,21, 3451.

Sensors2021,21, 3451 2 of 18 perception of fatigue, caused by homeostatic and psychological factors, and performance fatigability, which is in uenced by central and peripheral factors. During running, the human body undergoes shocks due to impacts with the ground. Performance fatigability can be assessed e.g., by means of changes in biomechanical quantities that are related to coping with such shocks. However, fatigue identi cation and management are commonly based solely on subjective estimates of fatigue that measure perception of fatigue. Subjective estimates of fatigue are very easy to use in practice, but they lack any assessment of performance fatigability. Inertial measurement units (IMUs) are non-intrusive sensors widely adopted to mea- sure biomechanical changes in human movement. IMUs can record biomechanical pa- rameters continuously, which can show changes due to physical fatigue [8,9]. Extensive research has been performed to detect biomechanical changes due to fatigue in running. Hip exion at initial contact was found to decrease between the start and the end of a fatiguing run [8,10]. Maximum knee exion angle can decrease [8] or increase [11] with fatigue depending on different running settings and subject characteristics. Maas et al. [12] showed that running experience could in uence knee exion, among other biomechani- cal parameters. Peak tibial (PTA) and peak sacral (PSA) accelerations are also recurring parameters studied in association with fatigue. Reenalda et al. [13] and Schutte et al. [14] found an increase in PTA due to fatigue, while Ruder et al. [15] found a decrease. Reenalda et al. further investigated shock attenuation between the tibia and the pelvis, nding an increase due to fatigue although both PTA and PSA increased as a consequence of fatigue [13]. Assessment of asymmetry in ankle, knee, and hip kinematics between a rested and fatigued state in running resulted in internal rotation of the knee showing the largest increase in asymmetry with fatigue [16]. Although signi cant changes in biomechanics have been repeatedly found when measuring running mechanics with IMUs, it is not clear yet whether these changes are suf cient to reliably detect fatigue over time in real-world applications. While fatigue detection in

fatigued state in running resulted in internal rotation of the knee showing the largest increase in asymmetry with fatigue [16]. Although signi cant changes in biomechanics have been repeatedly found when measuring running mechanics with IMUs, it is not clear yet whether these changes are suf cient to reliably detect fatigue over time in real-world applications. While fatigue detection in running has been based on non-automized detection of changes in biomechanical parameters, machine learning algorithms could have the bene t of rapid and easy application to identify fatigue. Machine learning algorithms could use as an input well-established biomechanical variables, as well as a wide range of statistical variables. Translation of biomechanical changes due to fatigue into machine learning fatigue detection algorithms has been performed in other elds. A clear example of such practice is in the area of industry work. Feeding a wide range of biomechanical parameters into a support vector machine classi cation algorithm led to a fatigue detection accuracy of 90% in working tasks [17]. Yet, few studies have focused on the detection of a fatigue condition in sports and running, especially in out-of-the-lab environments. Gholami et al. used machine learning techniques to detect the perceived exertion of runners on a treadmill using textile wearable sensors and assessed the importance of each sensor location, with the hip contributing more than the knee and the ankle to the nal coef cient of determination of 0.96 [18]. Buckley et al. located IMUs at the shanks and lumbar spine and compared three different locations and various machine learning classi ers to detect fatigue in outdoor running, obtaining a 75% accuracy with a single IMU placed at the lumbar spine [19]. While minimal sensor setups present the unequivocal advantage of being easy to wear, they might be missing substantial biomechanical information to improve fatigue detection accuracy. Here we aimed to assess the optimal combination of IMU locations at the lower limbs and trunk to detect fatigue levels in an outdoor run with a machine learning classi cation algorithm. We segmented IMU data into gait cycles and extracted biomechanical and statistical features, labeling

to wear, they might be missing substantial biomechanical information to improve fatigue detection accuracy. Here we aimed to assess the optimal combination of IMU locations at the lower limbs and trunk to detect fatigue levels in an outdoor run with a machine learning classi cation algorithm. We segmented IMU data into gait cycles and extracted biomechanical and statistical features, labeling data points with fatigue levels identi ed by means of subjective assessment of fatigue and heart rate (HR). IMU combinations of interest were selected and their fatigue detection performance was compared. It was hypothesized that larger biomechanical changes reported in the literature such as peak tibial accelerations would re ect in the combinations of sensor with higher fatigue detection accuracy. However, we

Sensors2021,21, 3451 3 of 18 expected statistical features derived from biomechanical quantities to also have a positive impact in the performance of the classi er. Findings of this study aim to assist translating current state-of-the-art knowledge of the biomechanical changes due to fatigue in running into detection of fatigue in real-world scenarios. 2. Materials and Methods The machine learning-based method that we implemented to detect fatigue in this study is summarized in Figure. Our work ow consisted of three main stages: data collection and processing (highlighted in yellow), development of the fatigue detection classi er (highlighted in blue) and performance evaluation of the classi er (highlighted in green). Each step in the work ow will be described in more detail in this section.Sensors 2021, 21, 3451 3 of 19 their fatigue detection performance was compared. It was hypothesized that larger bio- mechanical changes reported in the literature such as peak tibial accelerations would re- flect in the combinations of sensor with higher fatigue detection accuracy. However, we expected statistical features derived from biomechanical quantities to also have a positive impact in the performance of the classifier. Findings of this study aim to assist translating current state-of-the-art knowledge of the biomechanical changes due to fatigue in running into detection of fatigue in real-world scenarios. 2. Materials and Methods The machine learning-based method that we implemented to detect fatigue in this study is summarized in Figure 1. Our workflow consisted of three main stages: data col- lection and processing (highlighted in yellow), development of the fatigue detection clas- sifier (highlighted in blue) and performance evaluation of the classifier (highlighted in green). Each step in the workflow will be described in more detail in this section. Figure 1. Machine learning fatigue detection algorithm workflow. 2.1. Experimental Design Eight healthy runners were recruited (3 males 5 females, 24.3 ± 1.0 years, 174.8 ± 9.5 cm, 71.1 ± 8.8 kg, Table 1). Inclusion criteria consisted of the absence of major injuries in the previous year and having run at least 10 km per week in the previous six months. The ex- perimental protocol was approved

workflow. 2.1. Experimental Design Eight healthy runners were recruited (3 males 5 females, 24.3 ± 1.0 years, 174.8 ± 9.5 cm, 71.1 ± 8.8 kg, Table 1). Inclusion criteria consisted of the absence of major injuries in the previous year and having run at least 10 km per week in the previous six months. The ex- perimental protocol was approved by the Medical Ethical Review Committee ‘CMO Arn- hem-Nijmegen’ and all participants signed an informed consent form prior to participation. Figure 1.Machine learning fatigue detection algorithm work ow. 2.1. Experimental Design Eight healthy runners were recruited (3 males 5 females, 24.3 1.0 years, 174.8 9.5 cm, 71.1 8.8 kg, Table ous year and having run at least 10 km per week in the previous six months. The experimental protocol was approved by the Medical Ethical Review Committee `CMO Arnhem-Nijmegen' and all participants signed an informed consent form prior to participation.

Sensors2021,21, 3451 4 of 18 Table 1.Runners characteristics. Subject Age (Years) Body Mass (kg) Height (cm) Speed (km/h) Running Experience (Years) Sex S001 25 69 182 13.0 5 M S002 24 55 164 10.5 3 F S003 23 69 167 9.1 9 F S004 24 64 168 9.6 7.5 F S005 25 78 174 9.4 2 F S006 26 77 187 11.6 1.5 M S007 23 75 169 9.9 1.5 F S008 24 82 188 11.6 6 M Mean( 1STD) 24.3 1.0 71.1 8.8 174.8 9.510.6 1.4 4.4 2.8 - Subjects underwent a fatiguing protocol consisting of three distinct consecutive runs: 1.The rst run consisted of a 4000 m run (ten laps of the athletic track) at a constant speed, determined as the average speed of the subject during the best performance in the previous year on a 5 to 10 km race; 2. The second run was performed according to a fatiguing protocol. The speed in this fatiguing protocol started at the same level of the rst run and increased progressively by 0.2 km/h every 100 m. Perceived fatigue was assessed by means of a Borg Rating of Perceived Exertion (RPE) Scale (min-max score 6–20) [20], asked to the runner every 100 m. The fatiguing protocol was terminated once the RPE was equal to 16 (RPE between `hard' and `very hard') or higher, or, if such requirement was not met, after 1200 m; 3. The third run consisted of a 1200 m run (three laps of the athletic track), in which speed was kept constant and equal to the rst 4000 m run. Speed was controlled throughout the whole experimental protocol using a cyclist, proceeding at constant speed approximately 2 m in front of the runner. Except for the fatiguing protocol, half of each run was performed in clockwise direction and the other half in counterclockwise direction, in a randomized fashion to eliminate the effect of running direction on the biomechanics of the left and right leg. 2.2. Measurement Setup Xsens MTx IMUs (Xsens Technologies B. V., Enschede, The Netherlands) were attached with kinesiotape to eight body locations of

fatiguing protocol, half of each run was performed in clockwise direction and the other half in counterclockwise direction, in a randomized fashion to eliminate the effect of running direction on the biomechanics of the left and right leg. 2.2. Measurement Setup Xsens MTx IMUs (Xsens Technologies B. V., Enschede, The Netherlands) were attached with kinesiotape to eight body locations of the runner throughout the whole running experiment: left and right foot, left and right tibia, left and right thigh, pelvis and sternum (Figure). Double-sided tape was also attached between the IMU and the skin to limit skin artefacts. 3D accelerometer range of the IMUs is 16 g, 3D angular velocity range is 1200 /s, sampling frequency is 240 Hz. Running speed and HR were recorded simultaneously using a GPS watch (Garmin Forerunner 210, Garmin, Wichita, KS, USA). The bicycle speed was measured with a bicycle computer (Sigma BC 16.16 STS, Sigma, Neustadt, Germany) and shown in real time to the cyclist on a display. 2.3. Data Acquisition MVN Analyze (v2019.2.1, Xsens Technologies B. V., Enschede, The Netherlands) was used for data acquisition. A Kalman lter fusing accelerometers, gyroscopes and magnetometers data were used to estimate joint angles (left and right ankle, left and right knee, left and right hip) and segmental accelerations and angular velocities (left and right foot, left and right tibia, left and right thigh, pelvis and sternum) together with a biomechanical model [21].

Sensors2021,21, 3451 5 of 18Sensors 2021, 21, 3451 5 of 19 Figure 2. Measurement setup. IMUs are placed at both feet (1, highlighted left foot), both tibias (2, highlighted right tibia), both thighs (3, highlighted left thigh), pelvis (4) and sternum (5). 2.3. Data Acquisition MVN Analyze (v2019.2.1, Xsens Technologies B. V., Enschede, The Netherlands) was used for data acquisition. A Kalman filter fusing accelerometers, gyroscopes and magne- tometers data were used to estimate joint angles (left and right ankle, left and right knee, left and right hip) and segmental accelerations and angular velocities (left and right foot, left and right tibia, left and right thigh, pelvis and sternum) together with a biomechanical model [21]. 2.4. Data Analysis Three segments were extracted from the runs: laps 2–4 from the first run, identified as no fatigue condition (RPE = 6.0 ± 0.0); laps 8–10 from the first run, identified as mild fatigue condition (RPE = 11.7 ± 2.0); laps 1–3 from the third run, identified as heavy fatigue condition (RPE = 14.2 ± 3.0). The second run served only as a fatiguing protocol and dif- fered in length per subject, therefore was not included in the analysis. As per Figure 3, mean RPEs increased throughout the first run, and decreased during the last run after the fatiguing protocol (although considerably higher than the RPEs pre-fatiguing run), while HR kept increasing throughout the runs. Figure 2. Measurement setup. IMUs are placed at both feet (1, highlighted left foot), both tibias (2, highlighted right tibia), both thighs (3, highlighted left thigh), pelvis (4) and sternum (5). 2.4. Data Analysis Three segments were extracted from the runs: laps 2–4 from the rst run, identi ed as no fatigue condition (RPE = 6.0 0.0); laps 8–10 from the rst run, identi ed as mild fatigue condition (RPE = 11.7 2.0); laps 1–3 from the third run, identi ed as heavy fatigue condition (RPE = 14.2 3.0). The second run served only as a fatiguing protocol and differed in length per subject, therefore was not included in the analysis. As per Figure, mean RPEs increased throughout

8–10 from the rst run, identi ed as mild fatigue condition (RPE = 11.7 2.0); laps 1–3 from the third run, identi ed as heavy fatigue condition (RPE = 14.2 3.0). The second run served only as a fatiguing protocol and differed in length per subject, therefore was not included in the analysis. As per Figure, mean RPEs increased throughout the rst run, and decreased during the last run after the fatiguing protocol (although considerably higher than the RPEs pre-fatiguing run), while HR kept increasing throughout the runs. 2.5. Data Processing MATLAB R2019a (The MathWorks Inc., Natick, MA, USA) was used for data process- ing. Running gait segmentation was performed based on the pelvis velocity. First, the start and the end of each run were detected with the zero-crossing of the pelvis velocity in the sagittal plane. Then, downward peaks in pelvis velocity were calculated by means of a peak detection algorithm [22]. Left and right initial contact timepoints were determined based on the right knee angle. Joint angles, segmental acceleration magnitudes (a= q a 2 x+a 2 y+a 2 z ) and angular velocities in all three dimensions were cut into gait cycles starting at each initial contact and normalized at 150 data points.

This study evaluates IMU configurations for detecting running-induced fatigue using machine learning.