A Reliable User-Independent Motion Intent Detection from Transient EMG Data for Shoulder Joint

Motion intent detection (MID) through transient surface electromyographic (sEMG) is becoming central for the development of real-time assistive and rehabilitative technologies. However, the adaptation of such myoelectric interfaces in a user-independent framework represents a cutting-edge problem that hampers the practical usage of sEMG based human-machine interfaces. In this study least square canonical correlation analysis (LS-CCA) was employed together with SVM classifier in order to solve a shoulder joint MID problem. A publicly available dataset comprising eight healthy subjects was employed and four shoulder movements were considered. The LS-CCA was used for computing a subject-independent feature space using training data and then the trained SVM model was tested on an unseen subject, eventually implementing a leave-one-subject-out validation. Two window lengths for feature extraction (50 and 150 ms) and three feature sets were also compared. Overall, the best results were obtained using the 50 ms window length (multiuser classification accuracy about 70%), without any significant difference among the three feature sets.