Explainable deep learning framework for lower limb activity classification using reconstructed sEMG signals in healthy and pathological subjects

Recognizing human activities from Surface Electromyography (sEMG) signals is fraught with challenges, including noise susceptibility and signal crosstalk. Addressing these, our study analyses sEMG data from 11 healthy and 11 pathological subjects across three distinct activities: sitting, standing, and walking. We introduce a pioneering preprocessing approach that integrates Bandpass Filtering, Wavelet Denoising, and Ensemble Empirical Mode Decomposition (EEMD) for signal enhancement. A significant aspect of our approach involves the reconstruction of signals by selecting the top 50% of Intrinsic Mode Functions (IMFs) based on entropy, Signal-to-Noise Ratio (SNR), correlation, and energy, effectively capturing the most informative features of the sEMG signals. To counterbalance dataset imbalances and facilitate robust feature extraction, we applied Adaptive Synthetic (ADASYN) sampling and segmented the data into 256 ms windows with a 25% overlap. Our Convolution Neural Network (CNN) achieves remarkable classification accuracies: for sitting, 99.4% in healthy and 99.2% in pathological subjects; for standing, 99.7% and 98.8% respectively; and for walking, 99.4% and 98.6%, respectively. Furthermore, the integration of Explainable AI (XAI) through Permutation Feature Importance (PFI) provides critical insights into the significant impact of muscle signals, particularly highlighting the Rectus Femoris (RF) muscle’s role in sitting with leg extension, BF muscles’s role in standing with flexion. This comprehensive and innovative methodology not only overcomes the inherent challenges of sEMG signal analysis but also enhances the interpretability and reliability of activity recognition, marking a significant advancement for personalized healthcare interventions. • Advanced preprocessing pipeline using Bandpass Filtering, Wavelet Denoising, and EEMD. • Reconstructed sEMG signal using best 50% of IMFs based on entropy, SNR, correlation, and energy to enhance key signal features and denoise the signal. • Implemented the Adaptive Synthetic (ADASYN) sampling method to balance the class distribution in medical datasets, enhancing model generalizability and reducing bias towards majority classes. • Developed a tailored Convolutional Neural Network architecture with parallel convolutional pathways for each sEMG channel, ensuring precise feature learning from each muscle’s signal for accurate activity recognition. • Integrated Permutation Feature Importance (PFI) to identify key muscle signals contributing to activity classification, improving the model’s interpretability and insight into muscular involvement in lower limb movements. • Achieved 99.5% average classification accuracy across sitting, standing, and walking activities for healthy subjects and 98.9%. average classification accuracy for pathological subjects outperforming all previous state-of-the-artwork.