Weighted Average Confidence Score-based Ensemble Classification to Mitigate the Effect of Time on Myoelectric Control

Pattern recognition-based myoelectric control offers enhanced functionality to upper-limb prostheses compared to conventional proportional or direct control. However, this increased functionality often accompanies unreliability, with performance deteriorating over time. In this study, we introduced a Weighted Average Confidence Score (WACS)-based ensemble classification method to bolster robustness against the time effect. In this approach, electromyography (EMG) signals are first classified by various classifiers, each producing a confidence score. These classifiers are then assigned optimal weights and aggregated to derive the final prediction. The method was assessed on a fifteen-day consecutive EMG dataset involving seven intact-limb subjects performing six hand motions. It was benchmarked against Linear Discriminant Analysis (LDA) and Long Short-Term Memory (LSTM) neural networks. The results indicated that the WACS-based ensemble classification method surpassed LDA in performance and offered comparable but faster results than LSTM. It achieved errors of 1.75 ± 1.29% in within-day and 8.60 ± 6.10% in between-day classifications, respectively. These outcomes suggest that ensemble learning can effectively mitigate performance degradation over time and warrants further investigation.