Automatically Revealing Multi-Domain EEG Signatures Related to Motor Imagery Using a Fully-Interpretable...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalographic (EEG) signals are crucial for understanding brain functions.
• Interpretable Convolutional Neural Networks (CNNs) offer data-driven EEG analysis but often lack multi-domain interpretability.
• Existing methods primarily analyze EEG features in a single domain, limiting comprehensive neural network applications.

Purpose of the Study:

• To propose a fully interpretable CNN capable of learning and interpreting EEG features across frequency, spatial, and temporal domains.
• To develop a novel approach that overcomes the limitations of single-domain analysis in current interpretable CNNs for EEG.
• To create a comprehensive tool for automatic extraction of salient EEG features for various brain states.

Main Methods:

• Developed a novel interpretable CNN architecture.
• The network learns optimal bandpass filters (generalized Gaussian functions) and channel combinations.
• The model also learns optimal time-sample recombination for comprehensive feature extraction.
• Tested the approach on motor imagery decoding tasks.

Main Results:

• The proposed CNN significantly outperformed state-of-the-art interpretable CNNs in motor imagery decoding.
• Achieved the highest degree of feature interpretability across frequency, spatial, and temporal domains.
• Network-derived features aligned with known neural correlates of motor imagery.

Conclusions:

• The developed interpretable CNN provides a powerful tool for multi-domain EEG analysis.
• It enables easy interpretation of learned features across frequency, spatial, and temporal domains.
• This approach facilitates a deeper understanding of brain states by automatically extracting salient EEG features.