Analysis and Optimization of Novel Post-processing Method for Myoelectric Pattern Recognition

This paper describes a novel post-processing step for an electromyogram (EMG) pattern classification algorithm that is used in the control of myoelectric prosthetic hands. Amputees often find it difficult to control multiple degrees of freedom, but increasing numbers of prosthetic hands have multiple degrees of freedom. In general, larger numbers of classes tend to reduce the classification accuracy, and artificial neural networks have been used in previous studies for EMG pattern classification. The proposed post-processing algorithm stores the temporal sequence of classifications from the EMG pattern classification algorithm, and then runs a second classification based on the sequential patterns. We compared the output accuracy before and after the post-processing step. In our experiment, we set the training time for the EMG pattern classification algorithm to 1 s for each class, and used three channels of surface EMG signals. We considered nine EMG pattern classes, and recorded the output every 10-20 ms. We then analyzed the proposed algorithm and found seven classes to be the number required for optimal performance. The overall accuracy of the proposed system was 90.0% for seven classes and 89.8% for nine classes. The classification accuracy improved by 12.5% when using seven classes, and by 21.2% when using nine classes. We believe that this level of classification accuracy and other elements (the number of EMG channels and the training time) are sufficient for practical use with prosthetic hands.