Linear Discriminant Analysis with Bayesian Risk Parameters for Myoelectric Control

The linear discriminant analysis (LDA) classifier remains a standard in myoelectric control due to its simplicity, ease of implementation, and robustness. Despite this, challenges associated with the temporal evolution of the myoelectric signal may require flexibility beyond the capabilities of standard LDA. Recently proposed approaches have leveraged more complex systems, such as adaptive window framing or temporal convolutional neural networks to incorporate temporal structure. In this work, we explore the potential of exploiting parameters inherent to the LDA, which is conventionally applied assuming static and equal prior probabilities and uniform cost functions, to improve myoelectric control. First, a cost-modified version of the LDA (cLDA) is introduced to better reflect the comparatively high cost of active errors. Second, an adaptive priors version of the LDA (pLDA) is introduced to reflect the changing prior probabilities of classes in the myoelectric signal time series. Results are compared against the standard LDA classifier using a novel dataset comprised of continuous class transitions. Although no significant differences were observed in total error, the proposed cLDA algorithm yielded significantly lower active error rates than the LDA alone. Furthermore, both the cLDA and pLDA classification schemes produced significantly lower instability than the LDA classifier alone, as measured by spurious changes in the output decision stream. This work lays the groundwork for future research on these flexible classification schemes including context dependent cost arrays, different methods of priors adaptation, and combinations of the proposed cLDA and pLDA frameworks.