Explainable Deep Learning Model for EMG-Based Finger Angle Estimation Using Attention.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Electromyography (EMG) is crucial for detecting muscle activity but struggles with precise hand motion estimation.
• Accurately predicting finger joint angles from EMG signals remains a significant challenge in human-computer interaction and prosthetics.

Purpose of the Study:

• To develop an explainable deep learning model for estimating 14 finger joint angles from forearm EMG signals.
• To demonstrate the model's ability to generalize from single-finger to complex, multi-finger motions.
• To utilize an attention mechanism for interpretability of the EMG-to-joint angle relationship.

Main Methods:

• Proposed an encoder-decoder deep learning network incorporating an attention mechanism.
• Trained the model using forearm EMG signals to predict 14 finger joint angles.
• Validated the model's generalization capabilities on complex, random finger motion data.

Main Results:

• The deep learning model successfully estimated 14 finger joint angles from forearm EMG signals.
• The model demonstrated generalization from single-finger training data to complex hand motions.
• Attention matrix analysis revealed explainable, non-linear relationships between EMG signals and joint angles, consistent with experimental observations.

Conclusions:

• An explainable deep learning model with an attention mechanism can accurately estimate finger joint angles from forearm EMG.
• The proposed model offers a generalized approach, adaptable to complex hand movements.
• This method enhances the interpretability of EMG-based motion prediction, aligning sensor activity with specific finger movements.