سامانه همایش ها

English

صفحه اصلی / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Goniometry and Electromyography Data Analysis for Knee Health Diagnosis using Machine Learning

نویسندگان :

Mohammad-Reza Sayyed Noorani¹ Zahra Mahmoudi Anzabi² Sara Sharifi³

1- University of Tabriz 2- University of Tabriz 3- University of Tabriz

کلمات کلیدی :

Knee Health Diagnosis،Machine Learning،Feature Extraction،Goniometry،Surface Electromyography

چکیده :

In this study, we employed the Sánchez dataset [1] comprising synchronized knee goniometric measurements and surface electromyography (sEMG) recordings from major knee flexor and extensor muscles to develop a machine learning-based classification system for knee joint health assessment. The dataset included biomechanical data from 11 healthy controls and 11 participants with diagnosed knee pathologies. Our analysis focused only on the data collected during walking trials. Accordingly, training data prepared through kinematic monitoring of knee joint angles and subsequent segmentation of complete gait cycles - from initial heel-strike through terminal swing phase. Thus, we compiled 48 datasets from healthy controls and 173 datasets from participants with knee abnormalities. Each dataset included synchronized sEMG signals from four major muscles (rectus femoris, biceps femoris long head, vastus medialis, and semitendinosus) along with knee goniometry data, all of them were captured through complete gait cycles. Here, various combinations of statistical, temporal, and wavelet features using SVM, LDA, and KNN classifiers for knee health assessment were evaluated. Goniometric data alone achieved the best index with 97.7% accuracy (LDA/SVM models) when incorporating at least one feature from each type. For sEMG signal combinations, optimal performance (93.8% accuracy with LDA) was obtained using solely semitendinosus muscle data with complete feature sets. Comparative analysis revealed wavelet features as the least effective individually, while combined feature sets consistently yielded superior results. The sEMG signals from other individual muscles or their various combinations, regardless of feature selection approach, consistently demonstrated inferior classification performance.