Improving the performance of a neural-machine interface for prosthetic legs using adaptive pattern classifiers

Pattern classification has been used for design of neural-machine interface (NMI) that identifies user intent. Our previous NMI based on electromyographic (EMG) signals and intrinsic mechanical feedback has shown great promise for neural control of artificial legs. In order to make this NMI practical, however, it is desired that classification algorithms can adapt to EMG pattern variations over time, caused by various physical and physiological changes. This study aimed to develop an adaptive pattern recognition framework in the NMI to improve the robustness of NMI performance over time. Two adaptive algorithms, i.e. entropy-based adaptation and Learning From Testing Data (LIFT) adaptation, were presented and compared to the NMI with non-adaptive classifiers. Support vector machine (SVM) was selected as the basic classifier. Gradual change of EMG signals was simulated over time on EMG data collected from four transfemoral (TF) amputees. The preliminary results showed that the NMI with adaptive classifiers produced more consistent performance over time than the classifier without adaptation. The results of this preliminary study indicate the potential of using adaptive classifiers to improve the NMI reliability for neural control of powered prosthetic legs.