EMG-Based Estimation of Limb Movement Using Deep Learning With Recurrent Convolutional Neural Networks.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Electromyogram (EMG) signals contain neural information for limb movement.
• Signal variability poses challenges for accurate myoelectric control.
• Existing methods like Support Vector Regression (SVR) and Convolutional Neural Networks (CNNs) have limitations.

Purpose of the Study:

• To propose a novel deep learning model for estimating kinematic information from multi-channel EMG signals.
• To enhance the accuracy and robustness of myoelectric control.
• To investigate the combined efficacy of CNNs and Recurrent Neural Networks (RNNs) for EMG decoding.

Main Methods:

• A deep learning architecture combining CNNs and RNNs was developed.
• EMG signals were transformed into time-frequency frames for model input.
• The model was trained using gradient descent and backpropagation.
• Performance was evaluated against SVR and CNNs in eight healthy subjects for simultaneous and proportional limb movement estimation.

Main Results:

• The proposed CNN-RNN model demonstrated higher estimation accuracy compared to SVR and CNNs alone.
• The model exhibited superior robustness over time.
• Combining CNNs and RNNs significantly improved performance over CNNs alone.
• The deep architecture showed promising results in EMG decoding.

Conclusions:

• The novel deep learning model offers a promising approach for advanced myoelectric control.
• The integration of CNNs and RNNs enhances EMG signal processing for kinematic estimation.
• Further optimization of network structures can lead to increased accuracy and robustness in prosthetic control.