Sprinting-induced fatigue significantly compromises neuromuscular performance and elevates the risk of injury. Accurate monitoring of muscle fatigue is essential for designing personalized training and rehabilitation strategies. This study presents a novel method that integrates hybrid feature extraction with interpretable machine learning to assess fatigue during dynamic contraction.

In this work, time–frequency methods, namely Stockwell Transform (S-transform), B-Distribution (BD), and Extended Modified B-Distribution (EMBD), were applied to distinguish dynamic muscle fatigue states. Surface electromyography (sEMG) signals were recorded from the lower limb muscles of 14 healthy collegiate athletes during sprinting. The non-fatigue, fatigue progression, and fatigue segments of the signals were preprocessed and analyzed using these methods. From each method, thirteen features were extracted, and prominent features were selected using Genetic Algorithm (GA) and Principal Component Analysis (PCA). Classification of fatigue states was performed using four machine learning algorithms: Decision Tree, Support Vector Machine (SVM), Random Forest, and Artificial Neural Network (ANN). Furthermore, spectral features such as mean frequency and median frequency were analyzed to compare fatigue across different muscles.

The results demonstrate that fatigue is characterized by a progressive decline in median frequency (MDF) and mean frequency (MNF) over time. The gastrocnemius lateral head exhibited the steepest decrease in both MDF and MNF, indicating a higher susceptibility to fatigue during dynamic contractions, followed by the gastrocnemius medialis head. Comparative analysis of spectral features across sEMG segments revealed that the transition from fatigue progression to established fatigue occurred more rapidly than the shift from non-fatigue to fatigue progression. Classifier performance evaluation showed that the Random Forest model achieved the highest accuracy of 96.62% using features selected by the Genetic Algorithm (GA), outperforming models trained on Principal Component Analysis (PCA)-selected features (90.79%) and all features combined (92.83%). In contrast, the Support Vector Machine (SVM) classifier recorded the lowest accuracy at 66%.

The proposed method effectively detects dynamic muscle fatigue and shows strong potential for integration into real-time fatigue monitoring for wearable systems.