Real-Time Continuous Assessment of Fatigue from Surface Electromyography with Deep Learning for Training Load Regulation in Elite Cyclists

Objective To compare linear regression and deep learning models for quantifying exercise-induced muscle fatigue from raw surface electromyography (sEMG) signals, and to identify effective approaches for accurately assessing fatigue progression during a 30-second all-out cycling sprint in elite cyclists, so as to support precise training load regulation and individualized fatigue management. Methods Fourteen elite track cyclists performed a 30-second all-out cycling sprint. Surface electromyography signals were recorded from four key lower limb muscles: rectus femoris, biceps femoris, tibialis anterior, and lateral gastrocnemius. Exercise-induced fatigue was quantified by the continuous decline in power output throughout the sprint. A deep learning model integrating convolutional neural networks (CNN), bidirectional long short-term memory (Bi-LSTM) networks, and an attention mechanism was developed to directly predict fatigue progression from raw sEMG data in an end-to-end manner. For comparison, a linear regression model was trained using eight handcrafted time- and frequency-domain EMG features: root mean square (RMS), median frequency (MF), mean power frequency (MPF), mean frequency (MNF), mean frequency deviation (MDF), spectral entropy (SE), fractal dimension (FD), and Lempel-Ziv complexity (LZC). Results The proposed deep learning model significantly outperformed all baseline models, achieving a coefficient of determination (R²) of 0.94 ± 0.02 and a mean absolute error (MAE) of 2.13 ± 0.32. Compared to the linear regression model, the deep learning approach improved R² by over 50% and reduced MAE by more than two-thirds. Conclusion This study demonstrates that an end-to-end deep learning framework can accurately and continuously track muscle fatigue directly from raw sEMG signals during high-intensity cycling. These findings highlight the superiority of deep learning over traditional feature-based linear models and provide a promising tool for real-time, individualized fatigue monitoring in elite sports performance.