Deep Learning Movement Intent Decoders Trained With Dataset Aggregation for Prosthetic Limb Control.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Technology

Background:

• Traditional methods for decoding movement intent from electromyograms (EMGs) often degrade over time.
• Conventional neural network training algorithms may struggle with performance outside of observed training data domains.

Purpose of the Study:

• To present and evaluate four decoding methods for volitional movement intent from intramuscular EMG signals.
• To mitigate performance degradation and improve decoder robustness for prosthetic limb control.

Main Methods:

• Four decoding methods were developed: polynomial Kalman filters (KFs), Multilayer Perceptron (MLP) networks, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks.
• Decoders were trained using the Dataset Aggregation (DAgger) algorithm, augmenting training data iteratively.
• Performance was evaluated using EMG datasets from individuals with transradial amputation through short-term and long-term analyses.

Main Results:

• Short-term analyses showed CNN and MLP decoders outperformed KF and LSTM decoders by up to 60% in normalized mean-square error.
• Long-term analyses (0-150 days) indicated CNN, MLP, and LSTM decoders significantly outperformed KF decoders.

Conclusions:

• MLP- and CNN-based decoders trained with DAgger demonstrate potential for more accurate and naturalistic prosthetic hand control.
• These advanced decoding approaches offer a substantial improvement in the quality of life for individuals with limb loss.