A permutation importance and ensemble learning based feature selection approach for muscular intent decoding

Muscle signals are indeterministic and contain huge inter-subject variations. The work proposes a subject-specific feature selection approach employing permutation importance-based weight calculation to identify different hand movements correctly. The performance of the proposed method is evaluated in terms of accuracy, F1 score, and computational time. The study finds that merely 25% of the features are enough to predict the movements using the ensemble-based classifier. The accuracy and F1 score increment are almost 3-5% with only 25% features. The feature reduction significantly reduces the training and validation time by almost 40% compared to the time taken for the whole feature group.