Deep Learning-Based Assessment of POMA Scores Through EMG Derived Correlation and Coherence Matrices

Stroke is a critical medical condition characterized by the sudden interruption of blood flow to the brain, often resulting in cellular damage or, in severe cases, tissue necrosis. The consequences of stroke vary significantly among individuals, ranging from mild functional impairments to profound disabilities. Hence, gait rehabilitation is a cornerstone of stroke recovery, aiming to restore mobility and improve functional outcomes. Moreover, the effectiveness of rehabilitation interventions is commonly assessed using clinical scales. Among these, the Performance-Oriented Mobility Assessment (POMA) is widely used to evaluate balance, gait performance, and fall risk in impaired patients. However, clinical scales are typically administered through qualitative evaluations during clinical examinations, relying heavily on the subjective judgment of clinicians. While these assessments are valuable, they may lack consistency and fail to capture the complexity of neuromuscular and biomechanical impairments comprehensively. To address these limitations, it is crucial to develop instrumentalized, automated, and objective methods for assessing gait performance. This study explores the use of a deep learning-based regression model to predict POMA scores from electromyography (EMG) data. EMG signals collected from fourteen different muscles were analyzed to compute correlation and coherence matrices, which serve as features for the predictive model. The model achieved correlations of 0.89 and 0.81 throughout the validation and testing stages, respectively. The Bland- Altmann plot revealed a mild systematic error in the model. Although preliminary, these results suggest the potential for developing models that could significantly aid clinicians by providing accurate assessments of motor deficits following a stroke.