Altered Neuromuscular Time-Frequency Organization During Sit-to-Walk in Parkinson's Disease: An Explainable...

Neuromuscular dysfunction during the sit-to-walk (STW) task in Parkinson's disease (PD) remains poorly understood. We aimed to analyze surface electromyography (sEMG) signals across STW phases using explainable machine learning (ML) and deep learning approaches in PD. Individuals with PD (n = 101) and healthy controls (n = 50) performed a standardized STW task while sEMG signals were recorded bilaterally from eight lower-limb muscles. These signals were preprocessed and segmented into three phases in the STW task. Feature-based ML models were compared with convolutional neural networks (CNN) trained on wavelet-transformed sEMG spectrograms. Explainable artificial intelligence methods identified physiologically interpretable neuromuscular patterns. STW phase 1 was the most sensitive interval for detecting PD-related neuromuscular abnormalities. Among the feature-based classifiers, the random forest model demonstrated the highest performance under five-fold cross-validation, with an accuracy of 77.5% and an area under the receiver operating characteristic curveof 0.795. SHapley Additive exPlanations analysis identified mean and median frequency during STW phase 1, left biceps femoris integrated sEMG amplitude, and left rectus femoris-biceps femoris short head co-contraction as major contributors to classification. CNN-based analyses revealed earlier and temporally concentrated activation patterns in individuals with PD, particularly in proximal muscles critical for momentum generation and postural stabilization. Overall, individuals with PD exhibited distinct temporal, spectral, and coordinative patterns of lower-limb neuromuscular activity during the STW transition. The findings support the potential of sEMG-based explainable AI for characterizing PD-associated neuromuscular patterns during functional transitions.