Decoding transient sEMG data for intent motion recognition in transhumeral amputees

Abstract

The use of surface electromyographic (sEMG) signals, alongside pattern recognition (PR) systems, is fundamental in the design and control of assistive technologies. Transient sEMG signal epochs at the early beginning of the movement provide important information for upper-limb intent of motion recognition. However, only few studies investigated the role of transient sEMG for myoelectric control architectures. Therefore, in this work, focus was given to transient sEMG signals of intact-limb (IL) subjects and transhumeral amputees (AMP), who performed a series of shoulder movements. The role of the window length for feature extraction was investigated by sub-windowing the transient epochs at 200, 150, 100, and 50 ms window length (WL). Gaussian kernel discriminant analysis (SRKDA) and support vector machine (SVM) were used for recognizing seven classes of motion at different hold-out percentage of training/testing data, i.e. 70%–30%, 60%–40% and 50%–50%. In all the latter conditions, the median classification accuracy and F1 score were greater than 80% for both IL and AMP groups when using SRKDA. Wilcoxon rank sum test was employed to verify possible differences between WL conditions. Although the latter did not show significant differences, 100 ms WL showed the best classification performances for both groups (classification accuracy greater than 90%, near that of a usable PR system). Results demonstrated that a reliable motion intent recognition of shoulder joint in transhumeral amputee patients can be obtained employing transient sEMG epochs. This can be used in a better design of myoelectric control architectures of assistive technologies, involving the upper-limb for clinical use.