Evaluation of Support Vector Machine and Random Forest Models for EMG-Based Knee Motion Phase Classification

Knee-related conditions, such as anterior cruciate ligament injuries and meniscal ruptures, significantly affect mobility and quality of life. Although range-of-motion exercises are important for rehabilitation, interpreting electromyography signals related to these knee motions remains challenging owing to the complexity and variability of muscle activation patterns. This study investigated the classification of knee motion phases (resting, holding, and flexion/extension) by using surface electromyography signals from the biceps femoris and semitendinosus muscles in subjects with and without knee issues. Electromyography signals were preprocessed by removing outliers using the median absolute deviation and Kalman filtering. The motion phases were classified using a Support Vector Machine and Random Forest models within a subject-independent leave-one-subject-out evaluation framework. For subjects without knee issues, Random Forest performed with higher classification accuracy than Support Vector Machine for both muscles, with significant differences confirmed by McNemar’s test (p < 0.05). For subjects with knee issues, both classifiers performed with comparable mean accuracies of approximately 90% for both muscles, with no statistically significant difference between models (p > 0.05), despite the Random Forest exhibiting slightly more temporally coherent time-series segmentation. The results demonstrate that electromyography-based motion phase classification is feasible under strict subject-independent evaluation. Random Forest showed greater robustness in healthy subjects, whereas both classifiers performed equivalently under pathological conditions, enhancing their potential application in intelligent rehabilitation and clinical motion assessment systems.