EEG-PRISM Interprets AI Predictions for Brain Activity

Foundation models are rapidly advancing AI capabilities for Electroencephalography (EEG) analysis. However, current explainable AI (XAI) techniques typically provide attribution scores in the raw time-channel input space, which often doesn't align with how clinicians intuitively understand EEG data. There's a critical need for a universal method that can translate these AI predictions into more physiologically meaningful domains without altering or retraining the core foundation model. EEG-PRISM addresses this need by leveraging linear transformations and established backpropagation rules. It maps the time-channel attribution scores into alternative domains, specifically the frequency domain via an invertible Discrete Fourier Transform (DFT) and the source domain using an approximately invertible EEG generative model. This allows for a "physiologically-grounded" interpretation of what the AI model is focusing on. Evaluations using both simulated and real-world data demonstrated EEG-PRISM's effectiveness. It achieved near-perfect spectral recovery and 69.2% spatial accuracy in simulations. In clinical applications, it correctly identified salient delta-theta activity in epilepsy and accurately localized seizure onset regions with 50% accuracy. For autism, it localized predictive delta-alpha biomarkers to frontal and temporal regions, consistent with prior medical research. This work significantly enhances the interpretability of EEG foundation models, enabling clinically relevant insights like event localization and biomarker identification.