Explainable artificial intelligence using wavelet-transformed surface electromyography images to detect Parkinson’s disease during sit-to-walk

Background Neuromuscular abnormalities associated with Parkinson’s disease (PD) during the sit-to-walk (STW) task remain poorly characterized. We analyzed surface electromyography (sEMG) signals across three STW phases: overall, sit-to-stand, and gait initiation. Feature-based machine learning (ML) models were compared with wavelet spectrogram-based convolutional neural networks (CNN), and explainable artificial intelligence methods were applied to improve interpretability. Methods A total of 102 individuals with PD and 50 age-matched healthy controls were recruited. All participants performed a standardized STW task using a wireless sEMG system. Eight sensors were placed bilaterally on the lower limbs. sEMG signals were preprocessed and segmented into three STW phases. Feature-based ML models were compared with wavelet spectrogram-based CNN, and gradient-weighted class activation mapping (Grad-CAM) visualization was used to identify interpretable digital biomarkers reflecting PD-specific neuromuscular abnormalities. Results Among the tested models, the Random Forest classifier achieved the highest classification accuracy (92.6%). SHapley Additive exPlanations (SHAP) analysis revealed that frequency-domain features and co-contraction indices of the rectus femoris (RF) and biceps femoris short head (BFs) during STW phase 1 were the strongest predictors of PD. CNN-based visualization further highlighted earlier and more concentrated activation peaks in individuals with PD, particularly in the tibialis anterior, RF, and BFs, suggesting impaired activation timing and maladaptive co-contraction. Conclusion These findings demonstrate that while feature-based ML approaches provided higher classification accuracy, CNN-based analysis offered complementary interpretability by revealing muscle- and phase-specific activation patterns. This combined approach contributes to a more comprehensive understanding of neuromuscular dysfunction in PD during functional mobility tasks. Trial registration: Not applicable.