Mitigating the Impact of Electrode Shift on Classification Performance in Electromyography Applications Using...

Electromyography (EMG) signals have diverse applications, ranging from prosthetic hands and assistive suits to rehabilitation devices. Nonetheless, their performance suffers from cross-subject generalization issues, electrode shifts, and daily variability. In a previous study, while transfer learning narrowed the classification performance gap to -1% in an eight-class scenario under electrode shift, they imposed the burden of additional data collection and re-training. To address this issue in real-time prediction, we investigated a sliding-window normalization (SWN) technique that merges z-score normalization with sliding-window processing to align the EMG amplitude across channels and mitigate the performance degradation caused by electrode displacement. We validated SWN using experimental data from a right-arm trajectory-tracking task involving three motion classes (rest, flexion, and extension of the elbow). Offline analysis revealed that SWN mitigated accuracy degradation to -1.0% without additional data for re-training or multi-condition training, a 6.6% improvement compared with the -7.6% baseline without normalization. The advantage of SWN is that it operates with data from a single electrode position for training, which eliminates both the collection of multi-position training data and the calibration of deep learning models before practical use in EMG applications. Moreover, combining SWN with multi-position training exceeded the classification accuracy of the no-shift condition by 2.4%.