Forearm Orientation and Muscle Force Invariant Feature Selection Method for Myoelectric Pattern Recognition

Abstract

Electromyogram (EMG) signal-based prosthetic hand can restore an amputee's missing functionalities, which requires a faithful electromyogram pattern recognition (EMG-PR) system. However, forearm orientation and muscle force variation make the EMG-PR system more complex, and the problem becomes more complicated when muscle force levels and forearm orientations arise simultaneously. The problems can be minimized using a more significant number of features or high-density surface EMG, but it increases design complexity and needs higher computational power. In this regard, we have proposed a feature selection method that selects both feature and channel simultaneously. The proposed feature selection method selects only 7 to 20 features among 162 features with comparable or better performance. In this study, these selected features achieve a significant improvement in the accuracy, sensitivity, specificity, precision, F1 score, and Matthew correlation coefficient (MCC) by 3.18% to 4.28%, 9.14% to 12.85%, 1.83% to 2.57%, 8.30% to 10.99%, 9.22% to 13.92%, and 0.11 to 0.15, respectively comparing with four existing feature selection methods. In this research, the proposed feature selection method achieves a forearm orientation and muscle force invariant F1 score of 91.46% for training the k-nearest neighbor (KNN) classifier with two orientations, wrist fully supinated (O1) and wrist fully pronated (O3), with a medium force level. We have also achieved an F1 score of 93.27% for training the KNN classifier with all orientations with a medium force level. So, the proposed feature selection method would be very much helpful for finding the least dimensional features and achieving improved EMG-PR performance with multiple limiting factors.

Feature Set to sEMG Classification Obtained With Fisher Score

“…When there is a high number of features and/or channels (sensors) to describe the signal, it is said that it has a high dimensionality, which is not always recommended when classifying, so there are algorithms for its reduction [6], [7], [8]. In this sense, there are two main methods: feature selection algorithms and feature reduction algorithms.…”

“…Filter methods have low computational costs. However, selected features do not achieve good classification performance as they can sometimes miss some critical assumptions about the underlying regression function linking input variables to the output [7], [8]. The best-known filters are Information Gain, Gain Ratio, Term Variance, Mutual Information, Gini Index, Laplacian Score, Relief-F, and Fisher Score, among others.…”