Artificial Neural-Network EMG Classifier for Hand Movements Prediction

The objective of the present study was to develop a myoelectric controller able to classify specific hand movements from few milliseconds window of the EMG signals, detected by a patient’s forearm muscles. Five hand functional tasks have been selected: i) pitching; ii) grasp an object; iii) grasping; iv) lateral grasp; v) wave. The EMG task-selection classifier is designed so that the system predicts the intention to perform a certain task from EMG signals measured in a 100 ms window after the EMG onset corresponding to the electrophysiological delay. The prediction algorithm is based on an unsupervised K-means clustering followed by a cascade of artificial neural networks (ANN). It receives as input the 5 EMG signals measured in the 100 ms window and produces the predicted movement as output. Classifier performance is measured in percentage of correct classifications. The overall cascade ANN classifier produced a performance of 93%±7%, and of 72%±8% respectively on training and testing datasets, comparable with results obtained in literature. The results obtained during the tests on healthy subjects has provided satisfactory results, and the choice of the filtering algorithm to detect movement onset, and artificial neural networks cascade for classification allows a real-time implementation of the controller, opening the way of a real-time movement predictor.