Multi-label EMG Classification of Isotonic Hand Movements: A Suitable Method for Robotic Prosthesis Control

Classification is a crucial task in processing of surface electromyography (sEMG) signals. The Multi-Label Classification (ML) has gained prominence for biomedical applications, especially in gesture recognition for robotic prosthetic hand control. Differently of traditional method for classification involving three or more classes (multi-class), where each sample has only one output label, in the ML method, each sample may be referred to more than one output label. This work presents a pilot study of classification strategy using multi-label for hand gesture recognition. The methodology used to design the proposed system was the problem-transformation approach. The labels that compound the classifiers were developed based on the relationships of prosthesis operation and the anatomical nature of recognized gestures. In this way, it is easier to perform the motors of a prosthetic hand. The sEMG signals were obtained through the commercial Myo Armband (Thalmic Labs), in which data from 7 subjects were acquired performing 5 hand gestures. For the ML approach, the feature set was compound by L-Scale (LS), Maximum Fractal Length (MFL), Willison Amplitude (WAMP), and Mean Square Root (MSR). Nine classifiers were used, and the best classifiers with a accuracy of 97 and 98% with k-Nearest Neighbor (KNN) and Support Vector Machine with Radial Basis Function Kernel (SVMRBF). Regarding the single-label, no significant differences were observed between the multi-label tests.