EEG-PRISM: Physiologically-Grounded Interpretability of Predictions by EEG Foundation Models

Foundation models need better explanations for clinical use: Current AI explainability methods for EEG analysis show attribution scores in time-channel format, which doesn't match how clinicians actually think about brain activity, creating a gap between what AI systems identify and what doctors can understand and trust.

EEG-PRISM translates AI decisions into clinically relevant language: The method converts AI model explanations into frequency domain (brain wave patterns like delta-theta rhythms) and spatial domain (brain regions), without requiring retraining of the underlying AI model—making it a practical add-on tool for existing systems.

The approach shows promising real-world accuracy: In epilepsy cases, EEG-PRISM correctly identified delta-theta activity as most important and localized seizure onset regions with 50% accuracy; in autism, it pinpointed predictive biomarkers in frontal and temporal regions consistent with established medical knowledge.

Physiologically-grounded interpretability improves responsible AI deployment: By explaining AI predictions in terms clinicians recognize (specific brain frequencies and locations rather than abstract input features), this method helps ensure AI recommendations can be properly validated, questioned, and integrated into clinical decision-making rather than treated as black boxes.

The method works across multiple AI models and explanation techniques: EEG-PRISM was tested with five different foundation models and four different explainability methods, suggesting it could become a universal tool for making various EEG AI systems more interpretable to medical professionals.