Optimizing EMG Pattern Recognition for Prosthetic Hand Control: A Comparative Analysis of Windowing, Filtering, Features, and Classifiers Using the Ninapro Dataset

Abstract

Electromyography (EMG) enables intuitive pros-thetic hand control through pattern recognition, yet optimizing signal quality, processing time, feature selection, and classifiers remains challenging. Using Ninapro DB1 (27 subjects, 52 movements), this research addresses these via four parts: general optimizations, combined analysis, individual analysis, and hybrid modeling. Key results: optimal window (500-250 ms, 69.12% accuracy, 15.76 s time), filtering (stimulus=restimulus, 75.31% accuracy), features (slope sign changes, wavelength, total power, time-frequency energy via statistical categorization/outliers/Friedman test, p=1.14×10−7), classifiers (KNN/Random Forest at 90-97%). The hybridmodel integrates these for 90.49% accuracy, offering a robust framework for real-time applications.