A Machine learning framework for calf muscle fatigue assessment using IMU sensors and EMG-Based labeling

Calf muscle fatigue detection is essential for sports performance monitoring, injury prevention, rehabilitation, and tasks involving prolonged standing or repetitive lower-limb movements. While surface Electromyography (sEMG) is commonly used for fatigue assessment, its sensitivity to noise, electrode placement, and motion artifacts limits its practicality for continuous monitoring.

This study proposes an IMU-based approach for calf muscle fatigue detection, using sEMG only during the training phase for generating accurate fatigue labels. EMG-derived indicators such as Root Mean Square and Median Frequency were compared with multiple IMU features to identify those most strongly aligned with EMG behavior. Participants performed fatigue-inducing calf exercises, and percentage deviations from baseline EMG values were used to assign fatigue levels. Using these labels, IMU-based classification models were developed, including Random Forest for binary classification and a feedforward neural network for multi-class analysis.

The Random Forest classifier achieved 90.53% accuracy for the right calf and 90.19% for the left in binary classification, while the neural network reached 81.42% and 84.67% in multi-class classification. Subject-wise three-class classification further improved accuracy to 92.9% and 91.96% using Random Forest. These results demonstrate that IMU-only models can reliably detect calf muscle fatigue and are suitable for applications in sports, clinical practice, and daily activities, providing continuous, non-invasive monitoring to optimize performance and prevent overuse injuries.