An Interpretability Framework for Convolutional Neural Network-Based Electroencephalography Analysis Discovers New...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning (DL) models, especially CNNs, show potential for automated epileptic seizure detection from EEG.
• The
• black-box
• nature of these models hinders clinical adoption due to the need for interpretability and validation.

Purpose of the Study:

• To develop a novel interpretability method for CNN-based seizure detection models.
• To uncover meaningful spatial and spectral EEG biomarkers for seizure detection.
• To bridge the gap between DL and clinical EEG analysis for improved epilepsy diagnosis and treatment.

Main Methods:

• A novel interpretability approach combining frequency- and spatial-domain analysis for global and local explanations.
• Task-specific design, neurophysiological grounding, and cross-framework validation are incorporated.
• Results are visualized as heatmap matrices of feature importance (5 frequency bands * 5 brain regions), with statistically validated features.

Main Results:

• The method was validated on three CNN architectures, revealing distinct frequency band and brain region utilization for seizure detection.
• Global interpretation showed the best model using complementary biomarkers across multiple frequency bands.
• Local interpretation captured dynamic spectral shifts during seizures, aligning with known neurophysiological mechanisms like thalamocortical interactions and default mode network suppression.

Conclusions:

• The proposed method provides a tool for validating DL models in seizure detection.
• It facilitates the discovery of novel electrophysiological signatures in epilepsy.
• The interpretability approach enhances trust and clinical adoption of DL in EEG analysis for epilepsy management.