SE-CNN integrated with transfer learning: Enhancing myoelectric pattern recognition amid electrode shift and damage interferences

• Published in: Biomedical signal processing and control Vol. 111; p. 108411
• Main Authors: Zhang, Yue, Bao, Zhenchen, Huang, Rixi, Yin, Xiangyu, He, Bingwei, Liu, Yuqing
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2026
• Subjects: Convolutional neural network(CNN); Electrode damage; Electrode shift; Surface electromyography(sEMG); Transfer learning; Convolutional neural network(CNN); Electrode shift; Transfer learning; Surface electromyography(sEMG); Electrode damage
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.108411

Pattern recognition models trained on low-density surface electromyography (sEMG) sensors are susceptible to signal quality degradation and source variability. This study addresses the critical challenge of reduced gesture recognition accuracy in armband-based sEMG systems caused by concurrent interference of electrode shift and damage. We propose a hybrid approach integrating a convolutional neural network (CNN), a squeeze-and-excitation (SE) attention block, and transfer learning (TL). Data from seven hand gestures performed by nine subjects under electrode shift/damage were analyzed. The SE-CNN TL model achieved accuracies of 96.32 ± 1.29% (shift only), 94.98 ± 3.82% (damage only), and 94.30 ± 1.51% (concurrent interference)—significantly outperforming conventional and deep learning benchmarks. Notably, the accuracy under concurrent interference represents the highest level reported to date. This method demonstrates universality against diverse interferences and establishes a new state-of-the-art for concurrent interference mitigation in low-density sEMG systems. Our framework provides a generalized solution for robustness enhancement in sEMG-based pattern recognition. [Display omitted] •Reaching 94.30 ± 1.51% accuracy under concurrent electrode shift and damage.•Broad applicability in both of electrode shift alone and electrode damage alone.•Fine-tuning with single gesture data rapidly boosts robustness.•Attention mechanisms prioritize critical sEMG patterns.•Advancing wearable reliability for conditions under interference in low-density sEMG.