Use of regularized discriminant analysis improves myoelectric hand movement classification

Abstract

Linear discriminant analysis (LDA) is the most commonly used classification method for movement intention decoding from myoelectric signals. In this work, we review the performance of various discriminant analysis variants on the task of hand motion classification. We demonstrate that optimal classification performance is achieved with regularized discriminant analysis (RDA), a method which generalizes various class-conditional Gaussian classifiers, including LDA, quadratic discriminant analysis (QDA), and Gaussian naive Bayes (GNB). The RDA method offers a continuum between these models via tuning two hyper-parameters which control the amount of regularization applied to the estimated covariance matrices. In this study, we performed a systematic classification performance comparison on four datasets. Hand motion was decoded from myoelectric and inertial data recorded from 60 able-bodied and 12 amputee subjects whilst they performed a range of 40 movements. We found that when the regularization parameters of the RDA classifier were carefully tuned via cross-validation, classification accuracy was statistically higher by a large margin as compared to any other discriminant analysis method (average improvement of 13.7% over LDA). Importantly, our findings were consistent across the able-bodied and amputee populations. This observation provides supporting evidence that our proposed methodology could improve the performance of pattern recognition-based myoelectric prostheses.