Wearable Accelerometer-Based Temporal Modeling of Fatigue Progression During 400 m Running

Athletes may be affected by fatigue due to repetitive and physiological motions when running, and detecting fatigue is crucial to preventing injury and improving performance. This study was conducted to develop a methodology for processing acceleration signals to classify progressive fatigue states during a staged 400-m running event. Acceleration sensors were placed on the right calf, left ankle, and lower spine of 50 competitive and regular male runners who completed running tasks primarily engaging the lower extremity muscles. This approach focused on analyzing kinematic data to assess fatigue during running. Borg’s Rating of Perceived Exertion (RPE) was used to quantitatively label athletes’ fatigue data (RPE ≥ 15 for fatigue; RPE ≤ 14 for non-fatigue), corroborated by concurrent heart rate and blood oxygen saturation (SpO2) telemetry. Acceleration signals were pre-processed using a fourth-order Butterworth filter. Notably, 22 features were extracted from the 2400 datasets every 25 m during the 400 m run. Under a strict subject-independent leave-subjects-out validation protocol, machine learning algorithms classified running mechanics into non-fatigue and fatigue states, with fine K-nearest neighbor, fine Gaussian support vector machine, and fine random forest achieving testing accuracies of 75.00% (F1-score: 73.45%), 88.75% (F1-score: 88.61%), and 77.50% (F1-score: 77.12%), respectively. The deep learning recurrent neural network achieved the highest testing accuracy of 91.25% (F1-score: 91.14%). While this post hoc computational methodology establishes a mathematical baseline for mapping progressive running exhaustion specifically under a staged 400 m protocol, the findings represent an initial proof of concept derived from a young male cohort using offline processing. Future prospective validation under matched-velocity conditions, across female athletes, diverse age groups, varied terrains, and embedded edge devices is required to transition this workflow into generalizable, real-time athletic screening systems.