Lower limb Locomotion Mode Recognition based on sEMG: A pilot study in healthy individuals and stroke patients

• Published in: Biomedical signal processing and control Vol. 111; p. 108475
• Main Authors: Zhang, Changhe, Yu, Zidong, Wang, Xiaoyun, Chen, Ze-Jian, Deng, Chao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2026
• Subjects: Bioelectrical signal denoising; Deep learning; Inter-subject prediction; Locomotion Mode Recognition; Surface electromyography; Transfer learning; Deep learning; Inter-subject prediction; Bioelectrical signal denoising; Surface electromyography; Locomotion Mode Recognition; Transfer learning
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2025.108475

Nowadays, research on lower limb Locomotion Mode Recognition (LMR) based on Surface Electromyography (sEMG) and deep learning is attracting increasing attention, especially in fields such as sports rehabilitation and exoskeletons. Typically, directly measured sEMG signals are prone to various noise interferences that severely affect the recognition performance, whereas existing methods are insufficient to effectively handle noise within the spectrum. Additionally, current studies pay insufficient attention to the recognition performance of models in inter-subject scenarios. Therefore, this study proposes a novel LMR method. Specifically, a sEMG denoising framework based on mode decomposition and signal reconstruction is designed to adaptively identify and eliminate noise hidden within its spectrum. Several generic deep learning models are constructed to achieve end-to-end LMR using denoised signals, including typical CNN, serial CNN-RNNs, and parallel CNN-Transformer. On this basis, a transfer learning technique called weight-decomposed low-rank adaptation is integrated to maximize the applicability of the model in inter-subject scenarios with minimal increase in inference cost. A multi-metric evaluation framework is established to systematically analyze the proposed method from various aspects such as denoising effect, model performance, and time cost. Effectiveness is verified using a benchmark dataset of lower limb motions with ten healthy subjects, and a custom dataset including healthy individuals and stroke patients. Experimental results indicate that the proposed method outperforms existing state-of-the-art methods across multiple recognition tasks in both types of individuals, showing promising potential in the development of human-robot coordinated rehabilitation exoskeletons.