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

When clinicians look at an EEG trace, they think in terms of rhythms, scalp topographies, and the underlying neural generators that produced them. Most explainable‑AI tools for deep learning models, however, highlight importance scores only in the raw time‑channel matrix, a representation that feels alien to bedside interpretation. This mismatch forces experts to mentally translate abstract heatmaps back into familiar neurophysiological concepts, slowing adoption and limiting trust.

What You Need to Know

EEG‑PRISM addresses this gap by providing a universal, model‑agnostic way to re‑express any attribution map produced by an EEG foundation model into domains that clinicians already use. Starting from the gradient‑based attribution (e.g., Integrated Gradients or SHAP) computed in the time‑channel space, the method applies two linear transformations. First, an invertible discrete Fourier transform (DFT) converts the signal into the frequency domain, yielding attribution per electrode and per frequency band. Second, an approximately invertible EEG generative structure—built from a lead‑field model and a realistic head‑volume conductor—maps the time‑channel attributions to cortical source space, giving a rough estimate of which neural regions contributed most to the model’s decision.

Because both transforms are linear (or approximated linear) and rely only on standard backpropagation rules, EEG‑PRISM does not require retraining or fine‑tuning of the underlying foundation model. The approach works with any architecture—CNNs, Transformers, or hybrid models—as long as the model accepts raw multichannel EEG as input. The method preserves the additive property of attributions, so the sum of importance scores remains consistent across domains, enabling direct comparison of how much a delta rhythm versus a frontal source contributed to a prediction.

Why It Matters

Interpretability that speaks the clinician’s language bridges the trust gap between AI developers and end‑users. When a model flags an abnormal pattern, being able to show that the decision stems from heightened theta activity over temporal electrodes—or from increased source strength in the hippocampus—provides a actionable hypothesis that can be validated with traditional neurophysiology. This transparency also aids model debugging: if attributions highlight implausible frequencies or sources, developers can quickly identify data leaks or architectural biases.

Beyond trust, physiologically grounded explanations open the door to regulatory acceptance. Many medical‑device frameworks require evidence that an algorithm’s reasoning aligns with known biomedical mechanisms. By delivering explanations in frequency and source domains, EEG‑PRISM supplies exactly the kind of mechanistic insight that auditors and ethics committees look for, facilitating smoother pathways to clinical deployment.

Key Details

• Uses the standard DFT matrix; transformation is exactly invertible, preserving all information.
• Source mapping employs a lead‑field based generative model; approximation error quantified via singular‑value analysis (<5% average reconstruction error on benchmark head models).
• Method adds negligible computational overhead (<2 ms per sample on a CPU) because it relies on pre‑computed linear operators.
• Validated on three public EEG foundations (EEG‑BERT, BrainTS, and EEG‑Transformer) across seizure detection, sleep staging, and mental‑task classification tasks.
• Open‑source implementation includes PyTorch wrappers and visualisation tools for time‑frequency and source plots.

What’s Next

Future work will focus on tightening the source‑domain inversion by incorporating individualized head models from MRI, extending the framework to multimodal neuroimaging (e.g., EEG‑fMRI), and conducting user studies with neurologists to quantify improvements in decision speed and confidence when explanations are presented in physiologically meaningful terms.

📌 Source: Arxiv Ml