Non-human primate epidural ECoG analysis using explainable deep learning technology

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Explainable AI (XAI) is crucial for understanding artificial intelligence models.
• Existing XAI studies in neuroscience often lack deep neuroscientific interpretation of extracted features.
• Applying neural networks to neuroscientific data requires methods for explaining high-dimensional features.

Purpose of the Study:

• To actively explain high-dimensional learning features in neurophysiological data using XAI.
• To compare novel XAI-derived features with established neuroscientific findings.
• To develop and validate a deep learning approach for analyzing electrocorticogram (ECoG) data during movement.

Main Methods:

• Designed a 3D deep neural network (DNN) classifier to analyze neurophysiological data.
• Utilized 3D class activation mapping (3D CAM) to visualize high-dimensional classification features.
• Applied the 3D DNN and 3D CAM to classify monkey ECoG data from unimanual and bimanual movement experiments.

Main Results:

• The 3D DNN achieved superior classification accuracy compared to 2D DNN.
• 3D CAM analysis unexpectedly showed high activation in ipsilateral motor/somatosensory cortex during unimanual movement.
• The system identified critical temporal information at movement onset/offset for classifying bimanual movements.

Conclusions:

• This study is the first to integrate high-dimensional spatial, spectral, and temporal neurophysiological information with deep learning for explainability.
• The findings suggest motor cortex signals contain information about both contralateral and ipsilateral movements.
• The developed XAI methods offer a promising tool for neuroscience and electrophysiology research.