EMG Gesture Classification & Regression

A Neural Signal and Signal Processing project (EPFL, team of four) decoding motor intent from surface electromyography (EMG) on the NinaPro dataset, toward control of robotic prostheses. From 10-channel EMG, signals are stratified by stimulus and repetition and smoothed with a moving-average filter (a ~25-sample window chosen to cut noise without distorting the bursts).

Six time-domain features are extracted per channel: Mean Absolute Value, Standard Deviation, Maximum Absolute Value, Waveform Length, Root Mean Square and Slope Sign Changes, capturing signal energy, complexity and transitions. A Random Forest classifier, tuned by grid search with stratified K-fold cross-validation (300 trees, max depth 10), reaches 91.67% test accuracy over 12 movement classes (ROC-AUC 0.996); Mutual Information and SHAP both rank MAV and RMS as the top features, and keeping only the ~20 best pushes accuracy to 95.83% while improving interpretability.

A cross-subject study over all 27 subjects then exposes the hard part: intra-subject accuracy averages 91.2% but inter-subject drops to 33.8%, quantifying how electrode placement and anatomy break generalization. A third part moves from classification to regression, predicting continuous joint angles to enable finer prosthetic control such as half-closed hand positions.