Deep Learning for Enhanced Prosthetic Control: Real-Time Motor Intent Decoding for Simultaneous Control of Artificial...

Area of Science:

• Rehabilitation Engineering
• Biomedical Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Advanced prosthetic devices require seamless integration with daily life.
• Decoding motor intent from electromyography (EMG) signals is crucial for intuitive prosthetic control.
• Shallow neural networks have limitations in capturing complex EMG signal patterns.

Purpose of the Study:

• To compare the performance of deep learning architectures against shallow networks for motor intent decoding.
• To evaluate the effectiveness of different neural network models in real-time prosthetic control.
• To assess the generalizability of deep learning models across different user groups.

Main Methods:

• Four neural network architectures were evaluated: shallow feedforward, deep feedforward, temporal convolutional network, and convolutional neural network with squeeze-and-excitation.
• Real-time, human-in-the-loop experiments were conducted with able-bodied participants and an individual with an amputation.
• Electromyography (EMG) signals were used to decode motor intent.

Main Results:

• Deep learning architectures significantly outperformed shallow networks in decoding motor intent.
• Representation learning within deep networks effectively extracted motor control information from EMG signals.
• Performance improvements using deep neural networks were consistent across both able-bodied and amputee participants.

Conclusions:

• Deep neural networks offer superior performance for prosthetic control compared to shallow networks.
• Enhanced motor intent decoding using deep learning can lead to more reliable and precise prosthetic functionality.
• This approach has the potential to significantly improve prosthetic capabilities and quality of life for individuals with amputations.