(Machine) Learning From IMU-derived Biomechanics To Detect Physical Fatigue In An Outdoor Run

Abstract

Physical fatigue is a recurrent problem in running, leading to increased risk of injuries and decreased performance. Identification and management of fatigue helps reducing such negative effects. Inertial measurement units (IMUs) are able to record biomechanical parameters continuously, which can show changes due to fatigue. PURPOSE: To assess the performance of a machine learning classifier trained on IMU-derived parameters to detect physical fatigue. METHODS: 8 runners (3 M 5F, 24.3 ± 1.0 years, 174.8 ± 9.5 cm, 71.1 ± 8.8 kg) ran 13 laps of 400 m on an athletic track at a constant speed (10.6 ± 1.4 km/h) with 8 IMUs (240 Hz) attached to their body (feet, tibias, thighs, pelvis and sternum). Speed choice was based on past running performances. Three segments were extracted from the run: laps 2-4 (no fatigue stage, RPE = 6.0 ± 0.0); laps 8-10 (mild fatigue stage, RPE = 11.7 ± 2.0); laps 11-13 (heavy fatigue stage, RPE = 14.2 ± 3.0), run directly after a fatiguing protocol that followed lap 10. A random forest classifier was trained to detect fatigue stages with selected features from the 400 m moving average of the IMU-derived accelerations, angular velocities and joint angles. A leave-one-subject-out validation was performed to assess the classifier performance. RESULTS: See Table 1. Overall accuracy is 0.87 ± 0.09 for the three fatigue stages. Considering mild and heavy fatigue as a single fatigue stage, accuracy increases to 0.97 ± 0.05 (Sensitivity = 0.97 ± 0.05, Specificity = 0.95 ± 0.07). CONCLUSIONS: Fatigue was detected at an appropriate level using a machine learning classifier trained on a 8 IMUs setup, displaying the potential of a real-world application to classify different fatigue stages with a relatively unobtrusive sensor setup. Runners would benefit from even less obtrusive sensor setups. Performance of popular minimal sensor setups (1 or 2 IMUs) should be compared to that of larger setups (6+ IMUs) to assess the benefits and drawbacks of both. Supported by H2020 GA #826304Table 1: Performance metrics for the three fatigue stages (Mean ± SD).