Multilevel Assessment of Exercise Fatigue Utilizing Multiple Attention and Convolution Network (MACNet) based on Surface Electromyography

Background: Assessment of exercise fatigue is crucial for enhancing work capacity and minimizing the risk of injury. Surface electromyography (sEMG) has been used to quantitatively assess exercise fatigue as a new technology in recent years. However, the currently available research primarily distinguishes between fatigue and non-fatigue states, offering limited and less robust findings in multilevel evaluations. Methods: This study proposes a multiple attention and convolution network (MACNet) for a three-level assessment of muscle fatigue based on sEMG. Under the designed 50% maximum voluntary contraction experimental paradigm, sEMG signals and rate of perceived exertion scale are collected from 48 subjects. MACNet is developed to assess sEMG fatigue, incorporating improved temporal attention based on sliding window, multiscale convolution, and channel-spatial attention. Finally, GradCAM visualization is used to verify the developed model's interpretation, exploring the effects of sEMG channels and time-domain characteristics on exercise fatigue. Results: The average classification F1-Score and accuracy of MACNet are 83.95% and 84.11% for subject-wise and 82.83% and 82.43% for cross-subject, respectively. The GradCAM visualization highlights the greater contribution of the flexor digitorum superficialis and flexor digitorum profundus in evaluating high fatigue, along with the varied impact of time-domain features on exercise fatigue assessment. Conclusion: MACNet achieves the highest average classification accuracy and F1-Score, significantly higher than other state-of-the-art methods like SVM, RF, MFFNet, TSCNN, LMDANet, Conformer and MSFEnet, enhancing the extraction of exercise fatigue insights from sEMG channels and time-domain features.