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 dual-stream attention network with Luenberger feedback (DSA-Luen) for more accurate surface electromyography (sEMG) to torque mapping in stroke rehabilitation, improving robustness against inconsistencies.
Area of Science:
- Biomedical Engineering
- Neurorehabilitation
- Signal Processing
Background:
- Mapping surface electromyography (sEMG) to multi-joint torque is crucial for active stroke rehabilitation.
- Inter-set variations in repetitive exercises create inconsistencies in sEMG patterns, challenging model robustness.
Purpose of the Study:
- To design and validate a robust deep learning framework, DSA-Luen, for accurate sEMG to multi-joint torque mapping in upper-limb rehabilitation.
- To address sEMG pattern inconsistency and improve noise resistance in dynamic rehabilitation settings.
Main Methods:
- Developed a dual-stream time-frequency attention network incorporating Luenberger observation feedback (DSA-Luen).
- Integrated Real-Time Fast Fourier Transform (RTFFT) for dynamic sEMG information capture and a Luen-Mamba module for output feedback correction.
- Validated the model on datasets from healthy subjects and stroke patients performing upper-limb movements.
Main Results:
- DSA-Luen significantly outperformed baseline models in cross-set and cross-movement evaluations, achieving NRMSE of 0.130 and R² of 0.703 (cross-set).
- In stroke patient trials, DSA-Luen demonstrated strong adaptability, maintaining a positive mean R² of 0.527 and improving performance over baselines.
- The method showed significant reductions in Normalized Root Mean Square Error (NRMSE) and increases in R² compared to attention baselines and Informer.
Conclusions:
- The DSA-Luen framework offers enhanced adaptability and robustness for sEMG to torque mapping under non-ideal conditions in active rehabilitation.
- This approach provides a transferable foundation for intent decoding in personalized stroke rehabilitation, reducing the need for frequent recalibration.