IEEE Trans Neural Syst Rehabil Eng. 2026 Aug 3;PP. doi: 10.1109/TNSRE.2026.3718919. Online ahead of print.
ABSTRACT
Accurate decoding of motor intent in electromyogram (EMG)-based pattern recognition systems heavily depends on effective feature extraction. Although numerous techniques have been introduced, their performance remains suboptimal. Many fail to utilize spatial dependencies across EMG channels and overlook the distributional shift between training and test data. To overcome these challenges, this study introduces an unsupervised feature extraction method. It employs Dynamic Mode Decomposition (DMD) to isolate meaningful spatiotemporal patterns associated with hand gestures, followed by enhanced spatial representation through the Sample Covariance Matrix (SCM). To address dataset drift, the method aligns test data within the same non-Euclidean space as the training data. Experimental results show a significant improvement (p < 0.05) in decoding 13 distinct hand and finger gestures, with average accuracies of 99.89 ± 0.28% for amputees and 99.90 ± 0.31% for non-disabled subjects. Importantly, the method retains over 97% accuracy even when the number of EMG channels is reduced from 24 to just 3, suggesting its potential across both dense and sparse sensor configurations. This may be useful for low-cost prosthetic designs, where reducing the number of electrodes can decrease hardware complexity and cost. Thus, the proposed technique shows promise for improving the performance, reliability, and adaptability of EMG-based control systems, with potential relevance to clinical and commercial applications after further real-time and clinical validation.
PMID:42546000 | DOI:10.1109/TNSRE.2026.3718919