A Transformer Framework Informed by Muscle Anatomy and Sequence-to-Sequence Translation for Continuous Joint...

Area of Science:

• Biomedical Engineering
• Robotics
• Neuroscience

Background:

• Accurate prediction of patient motion intentions is crucial for developing effective assist-as-needed (AAN) control in rehabilitation robots.
• Existing methods often struggle with real-time prediction accuracy and generalizability.

Purpose of the Study:

• To redefine motion intention prediction as a sequence-to-sequence translation task.
• To develop and validate a novel 3DCNN-TF model for translating sEMG signals into kinematic representations.

Main Methods:

• The proposed 3DCNN-TF model utilizes a Transformer architecture for sequence-to-sequence translation.
• It incorporates a 3D Convolutional Neural Network (3DCNN) module to extract muscle synergy features from sEMG data.
• sEMG sliding windows are compiled into 'sentences' for input, and joint angles are generated autoregressively.

Main Results:

• The 3DCNN-TF model demonstrated superior performance in predicting wrist and knee joint angles compared to eight baseline models.
• Achieved high accuracy with average nRMSE of 6.2% (wrist) and 5.5% (knee), and R² of 95.5% (wrist) and 96.2% (knee).
• The model offers robust generalizability, computational efficiency ( < 2 min training), and can predict up to 300 ms in advance.

Conclusions:

• The 3DCNN-TF model represents a significant advancement in predicting patient motion intentions for AAN control.
• Its ability to accurately translate neural signals into kinematic representations enhances the real-time adaptability of rehabilitation robots.
• The model's efficiency, robustness, and predictive advance are critical for practical clinical applications.