Myoelectric pattern recognition of hand motions for stroke rehabilitation

Stroke is the fourth most common cause of death and can lead complex and long-term disability. In this regard, robotic-based rehabilitation could be an alternative for motion recovery. In this research we study how myoelectric signals (EMG) could be used to identify the fingers/hand motion through pattern recognition techniques. To this purpose, we implemented an experimental protocol on three subject groups: (I) non-stroke without hand impairments, (II) stroke without hand impairments and (III) stroke with hand impairments. The subjects performed a set of hand therapies to improve the range of motion and dexterity. Several methods for feature extraction, ranking and classification from EMG signals were implemented and the performance in the motion identification was compared. Specifically, three ranking methods: Two-sample T-test with feature variances, Separability Index, and the Davies-Boulding Index were used to determine the relevance of the features. As a result, dimensionality reduction was achieved by selecting only 50 features out of 136 with a comparable performance. Also, we compared three different classifiers: LDA, KNN and SVM. On average, the KNN classifier obtained a performance of 0.87 followed by the SVM with 0.82 and LDA with 0.74. Experimental results showed that we are able to identify the hand movements from subjects with a stroke event (group III) with 0.85 of correct classification rate average, which seems a promising approach in robotic-based rehabilitation assistance.