Quantifying Covariate Shift and Improving Electromyography Driven Gesture Recognition with Calibration and Sample Selection

Due to its non-stationary nature, surface electromyography (sEMG)-driven gesture recognition requires calibration to achieve satisfactory performance. Supervised domain adaptation is a commonly employed calibration technique. This technique involves combining one newly recorded repetition named calibration data of each gesture with all previously recorded data for each gesture under different conditions. This paper first quantified intra- and inter-conditions covariate shifts using the recognition error and Repeatability Index (RI), along with two newly introduced metrics: the non-overlapping index and the non-overlapping cluster count. Secondly, we extended the linear discriminant analysis (LDA) with the domain adaptation method by implementing a sample selection approach based on the repeatability index. We evaluated proposed techniques using two public datasets: the Multiday Ninapro DB6 and the Multiple Limb Position. Our domain adaptation approach with sample selection consistently outperformed the baseline model trained solely with calibration data in both datasets, showing improvements ranging from $0.7 \%$ to $3 \%$. Compared to the domain adaptation without sample selection, our method demonstrated a $0.3 \%$ performance increase for the Multiple Limb Position dataset and achieved comparable results for the Ninapro DB6, underscoring the efficacy of RI-based sample selection in the domain adaptation.