Control of Prosthetic Device Using Support Vector Machine Signal Classification Technique

Abstract

An appropriate classification of the surface myoelectric signals (MES) allows people with disabilities to control assistive prosthetic devices. The performance of these pattern recognition methods significantly affects the accuracy and smoothness of the target movements. We designed an intelligent Support Vector Machine (SVM) classifier to incorporate potential variations in electrode placement, thus achieving high accuracy for predictive control. MES from seven locations of the forearm were recorded over six different sessions. Despite meticulous attempt to keep the recording locations consistent between trials, slight shifts may still occur affecting the classification performance. We hypothesize that the machine learning algorithm is able to compensate for these variations. The recorded data was first processed using Discrete Wavelet Transform over 9 frequency bands. As a result, a 63-dimension embedding of the wavelet coefficients were used as the training data for the SVM classifiers. For each session of recordings, a new classifier was trained using only the data sets from the previous sessions. The new classifier was then tested with the data obtained in the current session. The performance of the classifier was evaluated by calculating the sensitivity and specificity. The result indicated that after a critical number of recording sessions, the classifier accuracy starts to reach a plateau, meaning that inclusions of new training data will not significant improve the performance of the classifier. It was observed that the effect of electrode placement variations was reduced and that the classification accuracy of >89% can be obtained.