Comparison of online adaptive learning algorithms for myoelectric hand control

Pattern recognition (PR) based myoelectric hand control has become a research focus in the field of rehabilitative engineer and intelligent control. However, the state of the art method is hardly adopted for clinical use because of signal interfered by shift, fatigue and user-unfriendly of retraining. The aim of this study is to evaluate the performance of different kinds of online algorithms in classifying the myoelectric hand motions, and reveal the key factors to classification accuracy of online learning algorithms. Two groups of experiments on intra-session and inter-session were designed to evaluate the classification and recognition performance of overall methods. The comparison results show that the second-order online learning algorithms outperformed the first-order algorithms in classification and recognition. Soft confidence-weighted learning performs best with 99% classification rate in same session and over 85% recognition rate in different session. This paper uncovers the online learning with large margin and confidence weight can always acquire a good property. In addition, online learning algorithms retrain the classification model by incorporating the testing data to the previous model by measuring the changes between the predicted label and true label which can improve the performance in long-term use.