Abstract
Background
Wearable lower-limb exoskeletons have the potential to support intention-driven control of lower-limb exoskeletons, but existing control strategies often rely on mechanical or manual triggers that fail to capture user intent.
Method
Subjects were recruited from 23/9/2024 to 10/2/2025. A BiLSTM detector was pretrained on a dataset collected from 50 healthy volunteers (45 for training, 5 for independent testing) using bilateral surface EMG recordings from six lower-limb muscles (12 EMG channels), 16-channel EEG, and hip–knee kinematics. Seven naive participants then completed ten 20-m outward-and-return walking trials (10 m outward and 10 m return) under each of three control modes (EMG, EEG, hybrid). Primary outcomes were triggering latency and classification accuracy. Triggering latency was defined as the time interval between the onset of the gait-transition event and the activation of the exoskeleton assistance command. This latency included the observation delay introduced by the sliding window, feature extraction time, BiLSTM inference time, and communication delay between the decoder and the exoskeleton controller. Usability was assessed with donning/doffing times and QUEST 2.0.
Results
The hybrid BiLSTM detector achieved higher classification accuracy (left: 91.3%; right: 86.6%) and shorter mean per-step triggering latency (left: 0.29 s; right: 0.28 s) than either EMG-only (mean 0.33 s) or EEG-only approaches (mean 0.30 s). Hybrid EEG–EMG fusion therefore improved decoding performance while reducing triggering latency compared with unimodal decoding strategies. The latency reduction relative to EMG-only control corresponded to a large effect size (Cohen’s d ≈ 0.88). Hybrid sessions also yielded shorter total session time (mean 32.1 min). Usability metrics demonstrated acceptable donning/doffing times and favorable QUEST 2.0 scores (mean 32.0/40).
Conclusion
These proof-of-concept results demonstrate that BiLSTM-based fusion of EEG and EMG improves responsiveness and classification reliability for exoskeleton assistance. These findings underscore the contribution of EEG-derived sensorimotor features and EMG information for intention-related gait transition detection within a multimodal real-time exoskeleton control framework. We discuss limitations related to sample size, artifact validation, and generalizability and identify next steps for patient studies and ergonomic optimization.