Related Experiment Videos
Generalization-Enhanced Cross-Set Upper-Limb Multi-Joint Torque Prediction via a Dual-Stream Time-Frequency Attention
Summary
This study introduces a novel network for mapping surface electromyography (sEMG) to multi-joint torque in stroke rehabilitation, improving accuracy despite inconsistent exercise data.
Area of Science:
- Biomedical Engineering
- Rehabilitation Robotics
- Neuroscience
Background:
- Mapping surface electromyography (sEMG) to multi-joint torque is crucial for effective active stroke rehabilitation.
- Inter-set variations in repetitive exercises introduce non-ideal factors, challenging sEMG pattern consistency and prediction accuracy.
Purpose of the Study:
- To design and validate a robust dual-stream time-frequency attention network with Luenberger observation feedback (DSA-Luen) for improved sEMG to torque mapping.
- To address sEMG pattern inconsistency caused by inter-set variations in upper-limb rehabilitation exercises.
Main Methods:
- Developed a DSA-Luen framework integrating a dual-stream attention network with Real-Time Fast Fourier Transform (RTFFT) and a Luen-Mamba module.
- Employed Luenberger observer-inspired output feedback for dynamic correction, noise suppression, and enhanced prediction accuracy.
- Validated the model on datasets from healthy subjects and stroke patients performing upper-limb movements.
Main Results:
- DSA-Luen significantly outperformed time-domain sequence models and attention baselines in cross-set and cross-movement evaluations.
- Achieved NRMSE of 0.130 and R² of 0.703 in cross-set evaluation, and NRMSE of 0.122 and R² of 0.704 in cross-movement evaluation.
- In stroke patient trials, DSA-Luen demonstrated strong adaptability and robustness, maintaining a positive mean R² of 0.527 and significantly outperforming baselines like Informer.
Conclusions:
- The DSA-Luen network offers a robust and adaptable solution for sEMG to multi-joint torque mapping in challenging rehabilitation scenarios.
- The proposed method provides a transferable foundation for intent decoding in personalized active rehabilitation, reducing the need for frequent recalibration.
- This approach enhances prediction accuracy and noise resistance, paving the way for more effective stroke rehabilitation technologies.