Action-Aware Multimodal Wavelet Fusion Network for Quantitative Elbow Motor Function Assessment Using sEMG and...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Artificial Intelligence in Medicine

Background:

• Accurate upper-limb motor assessment is crucial for stroke rehabilitation.
• Current clinical scales for Fugl-Meyer Assessment (FMA-UE) are subjective.
• Objective quantification methods are needed to improve rehabilitation outcomes.

Purpose of the Study:

• To develop an automated system for objective upper-limb motor assessment aligned with the FMA-UE.
• To integrate surface electromyography (sEMG) and robotic kinematics data for enhanced assessment.
• To create a computationally efficient model for real-time rehabilitation applications.

Main Methods:

• Proposed the Action-Aware Multimodal Wavelet Fusion Network (AMWFNet).
• Utilized Continuous Wavelet Transform (CWT) for signal processing.
• Implemented a modality gating mechanism and action embeddings for feature fusion.
• Validated the model on 40 participants (20 post-stroke, 20 healthy) across 18 reaching tasks.

Main Results:

• AMWFNet achieved 94.68% accuracy in six-class classification, outperforming baseline models (Random Forest, SVM, 1D-CNN) by over 9%.
• The model demonstrated superior performance compared to existing methods.
• The lightweight architecture allows for real-time integration into rehabilitation robots.

Conclusions:

• AMWFNet offers an objective and efficient solution for upper-limb motor assessment in stroke rehabilitation.
• The multimodal approach enhances quantification accuracy and provides actionable insights.
• Real-time capabilities facilitate seamless integration of assessment and training in robotic systems.