A NMF-based non-Euclidean adaptive feature extraction scheme for limb motion pattern decoding in pattern recognition system

Feature extraction is a crucial step in electromyogram (EMG)-based pattern recognition systems for decoding motor intents. However, despite the existence of numerous proposed techniques for feature extraction, their decoding performances have remained relatively low. Furthermore, these techniques are often evaluated without taking into account the drift between the training and test datasets. This study proposes a feature extraction scheme that operates in an unsupervised manner to address these limitations. This approach focuses on reducing drift between the training and test sets by utilizing feature adaptation based on non-negative matrix factorization (NMF) and Riemann operations. Additionally, we minimize drift by aligning the distribution of the test data with that of the training set. The results demonstrate that the proposed feature extraction technique exhibits significantly higher performance (p < 0.05) in decoding motor intent for 13 hand and finger movements, achieving an average accuracy of 99.91 ± 0.35% for amputee participants and 99.99 ± 0.02% for able-bodied participants. We also conducted further investigations to assess the effectiveness of the proposed feature scheme against varied signal-to-noise ratios (SNRs). These investigations revealed that our technique outperforms other feature extraction techniques in terms of decoding performance, even in the presence of varied SNRs. Overall, the findings show that the proposed feature extraction technique can effectively enhance the reliability and robustness of EMG control systems in both clinical and commercial applications.