Exercise ECG classification based on HRV features induced by robust R-peak detection model

Exercise-induced fatigue assessment via ECG classification relies on accurate R-peak detection for reliable HRV features. Addressing the lack of robust models for noisy exercise ECGs, we propose UNet-M-D, integrating positional encoding, multi-head self-attention, and dynamic convolution. Evaluated on GUDB and EPFL datasets, it achieves superior R-peak detection performance (up to 99.2% accuracy) with high noise resilience (6-18 SNR). Using optimally selected HRV features, our method attains 77.4% accuracy in fatigue classification, providing a scientific basis for sports health management and training adjustment.