Enhancing generalizability in classification of peripheral neural recordings with graph neural network

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Peripheral nervous system (PNS) signals are vital for biological communication.
• Decoding neural signals from PNS recordings is challenging due to complex spatiotemporal patterns.
• Existing methods struggle to fully capture spatial information from nerve cuff electrodes.

Purpose of the Study:

• To develop a graph-based learning approach for improved neural signal classification.
• To incorporate the physical geometry of nerve cuff electrodes into the decoding model.
• To enhance the decoding performance of neural signals from peripheral nerve recordings.

Main Methods:

• Utilized a publicly available dataset of neural recordings from eight rats using a 56-channel nerve cuff electrode.
• Constructed graphs where nodes represent electrode time series and edges represent geodesic distances between electrodes.
• Employed leave-one-out cross-validation for generalizability and within-rat testing for performance evaluation.

Main Results:

• Achieved a mean F1 score of 65.03% in generalizability evaluation, a 17.74% improvement over previous studies.
• Attained a mean F1 score of 77.50% in within-rat testing, a 3.14% increase.
• Demonstrated that incorporating recording geometry significantly improves decoding performance, especially with limited data.

Conclusions:

• Graph-based learning effectively captures spatio-temporal information for neural signal classification.
• Integrating physical recording geometry is crucial for enhancing decoding accuracy in PNS recordings.
• This approach offers a promising method for improving communication between biological systems via neural decoding.