Deep Learning-Based Approaches for Decoding Motor Intent From Peripheral Nerve Signals.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Deep learning effectively decodes motor intent from neural signals.
• Deep neural networks are often computationally complex, hindering real-time application.
• Developing efficient deep learning for real-time motor decoding is crucial for clinical translation.

Purpose of the Study:

• To investigate methods for enhancing the efficiency of deep learning-based motor decoding for real-time implementation.
• To analyze the trade-offs between different feature extraction and model deployment strategies.
• To inform the future development of accurate, low-latency motor decoders for clinical use.

Main Methods:

• Recorded neural data from amputees' residual peripheral nerves.
• Applied feature extraction techniques to reduce data dimensionality.
• Investigated one-step (1S) and two-step (2S) deep learning model deployment strategies.
• Predicted individual finger movements and combinations using recurrent neural networks (RNNs) and machine learning algorithms.

Main Results:

• The one-step (1S) approach with RNNs showed superior prediction accuracy on large datasets.
• The two-step (2S) approach, using classification before trajectory prediction, enabled comparable decoding with limited data.
• Both machine learning and deep learning achieved high accuracy (0.99) and F1 scores in the classification stage.
• The 2S approach resulted in comparable regression performance (MSE, VAF) to the 1S approach.

Conclusions:

• Feature extraction and a two-step approach can significantly improve the efficiency of deep learning motor decoding.
• Machine learning offers a simpler implementation for the two-step approach with comparable outcomes to deep learning.
• The findings provide a roadmap for implementing real-time, high-accuracy deep learning motor decoders in clinical settings.