Analyzing Frequency-Space-Time EEG Signatures via Interpretable Neural Networks: A Simulation Study

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Event-related electroencephalography (EEG) analysis is crucial for understanding brain function.
• Traditional methods require extensive preprocessing and may limit reproducibility.
• Task-relevant neural activity can be obscured by current analytical limitations.

Purpose of the Study:

• To validate an interpretable convolutional neural network (CNN) for EEG analysis.
• To automatically highlight frequency, spatial, and temporal EEG signatures in a data-driven manner.
• To provide an end-to-end, reproducible method for characterizing brain functions.

Main Methods:

• Simulated single-trial EEG data with imposed modulations across visual oddball and motor tasks.
• Applied an interpretable CNN at the single-participant level to analyze simulated EEG data.
• Quantitatively compared CNN-derived signatures with ground-truth simulations using localization errors and accuracies.

Main Results:

• The CNN successfully reproduced imposed spectral, spatial, and temporal modulations in simulated EEG.
• Achieved high localization accuracies: up to 85.3% spectral, 97.8% spatial, and 97.1% temporal.
• Demonstrated low localization errors within established EEG resolution limits.

Conclusions:

• The interpretable CNN accurately recovered task-relevant EEG signatures across multiple domains.
• Validated the use of CNNs for robust and reproducible EEG analysis.
• Establishes a foundation for trustworthy, participant-specific deep learning tools for EEG interpretation.