PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Deeksha M Shama, Punnisa Amornsirikul, Archana Venkataraman

arXiv:2608.13676v1cs.LG

TL;DR

Clinical interpretations of EEG foundation models remain difficult because post-hoc attributions are expressed in time-channel space rather than physiologically meaningful spectral and spatial domains. EEG-PRISM maps those attributions without modifying or retraining the model, recovering relevant frequency patterns and spatial regions across simulated and real-world analyses.

  • Problem

    Post-hoc EEG explanations remain difficult to interpret clinically because time-channel attributions do not directly represent spectral and spatial phenomena.

  • Method

    EEG-PRISM uses attribution propagation and linear transformations to project foundation-model attributions into spectral and spatial domains without modifying or retraining the model.

  • Results

    Across five foundation models and four attribution methods, EEG-PRISM recovered relevant frequency patterns and spatial regions in simulated and real-world epilepsy and autism analyses.

  • Takeaways & Limitations

    EEG-PRISM provides a model-agnostic post-hoc route to more clinically grounded interpretations of EEG foundation-model predictions.

  • Takeaways & Limitations

    EEG-PRISM analyzes spectral and source domains separately, and its theoretical guarantees currently cover only linear mappings.

Abstract

from arXiv · show

Objective: Foundation models represent the next advancement in AI for EEG analysis; however current explainable AI techniques provide attribution scores in the time-channel input space, which is mismatched to clinical intuition about EEG. Thus, there is a critical need for a universal method that can extend the interpretability of any foundation model to alternative and physiologically relevant domains without modifying or retraining the underlying model. Methods: EEG-PRISM leverages linear transformations and established backpropagation rules to map time-channel attribution scores into alternative domains. We derive mappings to the frequency domain via an invertible DFT and to the source domain via an approximately invertible EEG generative model. We evaluate EEG-PRISM in simulated and real data, assessing recovery of ground-truth phenomena across domains with five foundation models and four AI explainers. Results: In simulation, EEG-PRISM achieves near-perfect spectral recovery and 69.2% spatial accuracy. In epilepsy, EEG-PRISM correctly determines that delta-theta activity is most salient and correctly localizes the seizure onset region with 50% accuracy. In autism, EEG-PRISM localizes the predictive delta-alpha biomarkers to frontal and temporal regions, consistent with prior work. Conclusion: EEG-PRISM is a theoretically-grounded post-hoc attribution method with accurate mapping into the spectral and spatial domains. It supports window-level analysis of transient events (e.g., seizures) and group-level identification of clinically relevant biomarkers (e.g., autism), thus advancing interpretable EEG foundation models. Significance: This work enables physiologically-grounded interpretation of EEG foundation models and supports clinically relevant insights such as event localization and biomarker identification.

I. INTRODUCTION · A. EEG Foundation Models as Black-Box Computational Tools · B. Explainability in Deep Learning Pipelines

EEG foundation models learn directly from time–channel signals and offer strong, adaptable prediction, but their black-box explanations remain misaligned with clinically meaningful spectral and spatial EEG patterns. Post-hoc attribution methods quantify input contributions without modifying the model, yet their time–channel outputs can be difficult for clinicians to interpret or act upon.

  • I. INTRODUCTION: EEG captures temporal dynamics and structured spatial and spectral activity linked to neurological disorders and pathology.These properties make EEG clinically informative while contributing to its interpretive complexity.
  • I. INTRODUCTION: EEG source-domain modeling represents scalp signals as distributed neural activity propagated through head tissues by a forward model.Inverting this model can approximately map scalp activity to cortical regions, including seizure-onset localization.
  • I. INTRODUCTION: EEG’s high dimensionality, continuous multichannel structure, and artifact contamination make manual interpretation challenging despite strong AI prediction accuracy.These challenges motivate learned representations that generalize across datasets and tasks.
  • A. EEG Foundation Models as Black-Box Computational Tools: Foundation models pretrain on thousands of recording hours and use self-supervised learning to adapt across downstream EEG prediction tasks.This setup extends earlier deep networks and supports flexible use across tasks.
  • A. EEG Foundation Models as Black-Box Computational Tools: Foundation models learn directly from time-channel inputs, enabling rapid adaptation and strong performance while leaving their decision mechanisms difficult to interpret.Current explainable AI methods quantify input-output relationships but do not provide explanations in clinically meaningful spectral and spatial domains.
  • B. Explainability in Deep Learning Pipelines: Post-hoc, model-agnostic explainers assign attribution scores to quantify each input feature’s contribution to a prediction without modifying the trained black-box model.They address why a particular prediction is made rather than providing full model transparency.
  • B. Explainability in Deep Learning Pipelines: Time–channel attributions from EEG foundation models may be clinically misaligned because millisecond-scale temporal and correlated channel patterns do not always represent actionable phenomena.Post-hoc explainers include LRP, Integrated Gradients, DeepLIFT, and SHAP/DeepSHAP, which use distinct backpropagation or attribution principles.

C. Our Contributions

EEG-PRISM is a mathematically grounded, model-agnostic post-hoc framework that maps EEG foundation-model attributions into clinically meaningful frequency and source domains. Its contributions include theoretical guarantees, validation across five foundation models and real biomedical tasks, and alignment with clinical and research practice.

  • Framework: EEG-PRISM maps EEG foundation-model attributions into clinically meaningful frequency and source domains using DFT and inverse modeling.The framework includes a derived error bound and guaranteed compatibility across multiple explainers via backpropagation.
  • Framework: EEG-PRISM is model-agnostic and post-hoc, requiring no modification to existing foundation models and adding minimal computational overhead.This design extends interpretability without changing the underlying model.
  • Validation: EEG-PRISM demonstrates attribution fidelity across five EEG foundation models in simulations with known spectral- and source-domain ground truth.The simulations evaluate recovery against known ground truth in both domains.
  • Validation: EEG-PRISM is validated in seizure localization and autism biomarker discovery, supporting subject-specific event detection and group-level neural characterization.These real-world biomedical applications test both event-level and group-level interpretation.
  • Overall contribution: Together, these contributions provide a tool that aligns AI-based EEG analysis with clinical and research practice.The paper presents EEG-PRISM as a valuable tool for this alignment.

II. METHODS · A. Mapping Attribution Scores to Clinically Relevant Domains

EEG-PRISM post-hoc maps foundation-model attributions from the EEG time–channel space into clinically meaningful spectral and spatial domains. Its theoretically grounded linear transformation preserves key explainer properties without modifying the underlying model.

  • II. METHODS: EEG-PRISM applies foundation-model explainers post-hoc to obtain attributions beyond the input time–channel space, including spectral and spatial domains.The overall pipeline supports downstream clinically relevant applications.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: For EEG input X ∈ R^C×T, the method transforms row- or column-space vectors x into target-subspace vectors z through a fixed linear relation x = Wz.The target dimension satisfies M ≥ N, with W ∈ R^N×M.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: Proposition 1 maps target-space attributions A_z to signal-weighted linear combinations of the corresponding input-space attributions A_x.The weights are determined by the linear transformation W.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: The mapping follows by composing the foundation model with the fixed linear transformation, allowing standard post-hoc attribution rules to apply.The proof treats the transformed model as G̃(z) := G(Wz) = G(x) = y.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: Gradient-based explainers propagate attributions through W via the chain rule, while Integrated Gradients additionally uses linearity of path-integral accumulation.The derivation covers Gradient times Input and Integrated Gradients.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: DeepLIFT and DeepSHAP redistribute attributions linearly under fixed affine transformations using backpropagation and discrete-differential chain rules.Setting the reference to zero makes the DeepLIFT relation equivalent to Proposition 1, while DeepSHAP averages over multiple references.
  • A. Mapping Attribution Scores to Clinically Relevant Domains: The resulting equation is theoretically grounded, preserves explainer properties such as faithfulness, conservation, and stability, and leaves G(·) unchanged.EEG-PRISM therefore maps time–channel attributions into spectral and spatial domains without modifying the foundation model.

1) Spectral Domain:

EEG-PRISM maps time-domain EEG attributions into the spectral domain using an invertible DFT-based transformation. It handles complex coefficients by separating real and imaginary components, then averages channel-level spectral attributions for an overall representation.

  • Spectral Domain:: The spectral representation is obtained independently for each EEG channel using the DFT, a predefined linear and invertible operation.This enables attribution mapping from the time domain without changing the underlying signal transformation.
  • Spectral Domain:: Because DFT coefficients are complex-valued, EEG-PRISM decomposes them into real and imaginary components before attribution mapping.The components are concatenated as [ Re(zc) | Im(zc) ], and the inverse DFT reconstructs the EEG data.
  • Spectral Domain:: The time-to-spectral attribution weights are cos(2πji/T) for the real part and −sin(2πji/T) for the imaginary part.These weights are derived from the linear transformation underlying the DFT mapping.
  • Spectral Domain:: Spectral attributions are computed by combining the attribution scores of the real and imaginary components, then averaging across electrode channels.The channel average yields the overall spectral-domain attribution AF for the EEG data.

2) Spatial Domain:

EEG-PRISM projects time-channel attributions into the spatial source domain using the EEG forward model and its linear transformation weights. The method preserves linear fidelity, while source-estimation errors remain governed by the inverse solver.

  • Spatial Domain: EEG-PRISM uses the linear EEG forward model, with W ∈ C×M derived from biophysically informed head-conductivity models, to map channel attributions to sources.The transformation weights W_ji are obtained directly from the model’s built-in forward matrix.
  • Spatial Domain: Instantaneous source-level attribution scores are computed after substituting the forward-model weights and are averaged across time to obtain source-domain attribution A_S.
  • Spatial Domain: Source activity estimation is an ill-posed inverse problem because cortical sources often outnumber EEG channels, requiring appropriate regularization.The source count satisfies M ≫ C.
  • Spatial Domain: EEG-PRISM propagates source-estimation errors linearly without additional distortion beyond a transformation-dependent constant scaling factor.The method operates post-hoc and does not amplify or introduce errors into the inverse mapping.

B. Datasets

The experiments use three datasets: one simulated dataset for quantitative evaluation against known ground truth and two real-world datasets demonstrating clinical potential.

  • The study includes three datasets: one simulated dataset and two real-world datasets.
  • The simulated dataset supports quantitative evaluation of EEG-PRISM against a known ground truth.
  • The two real-world datasets demonstrate EEG-PRISM’s clinical potential.

1) Simulated Dataset:

The simulated dataset uses SEREEGA to generate EEG with source-to-scalp forward modeling, enabling controlled evaluation of EEG-PRISM across source-linked classes and settings.

  • Simulated Dataset: The dataset uses SEREEGA with a 32-channel Biosemi montage and the ICBM New York head model containing 74,382 sources.Classes are explicitly tied to underlying source-level signal characteristics for controlled evaluation across different settings.

2) TUSZ Dataset: · 3) ACE Dataset: · C. Experimental Setup

The study evaluates EEG-PRISM on focal-seizure data from TUSZ and resting-state autism data from ACE. Experiments adapt five EEG foundation models, apply four attribution explainers, and map time-channel scores into spectral and source domains.

  • 2) TUSZ Dataset:: TUSZ includes 124 subjects with focal seizure onsets and 19-channel average-referenced EEG recordings annotated by experts.Subjects were 58 male and 66 female, aged 19–91 years; recordings contributed randomly cropped non-seizure and seizure data.
  • 2) TUSZ Dataset:: TUSZ models classify seizure versus non-seizure windows, while clinical notes provide seizure-onset channels and relevant frequency bands.Patient-level onsets comprise 31 left frontal, 14 right frontal, 26 left posterior, 45 right posterior, and 8 central cases.
  • 3) ACE Dataset:: ACE contributes 177 participants meeting quality and assessment criteria, including 87 autistic participants and 90 typically developing controls.Data were acquired with a 128-channel EGI Net Amps 300 system using HydroCel nets.
  • 3) ACE Dataset:: ACE models perform subject-level autism-versus-control classification using each participant’s available time-channel EEG.EEG-PRISM spectral and spatial attribution maps are evaluated against findings in the autism-spectrum-disorder literature.
  • Preprocessing:: All datasets use uniform preprocessing: signals are downsampled to 200 Hz and segmented into 5-second simulated or 2-second real-world windows.The real-world window size improves seizure-detection resolution and matches the ACE protocol.
  • C. Experimental Setup: The experimental workflow trains foundation models, computes input attribution scores, and maps them into spectral and spatial domains.EEG-PRISM is applied to EEGNet, AttentionBaseNet, LaBraM, CBramod, and LUNA; the first two are trained from scratch, while the others are fine-tuned from released weights.
  • C. Experimental Setup: Models receive task-specific classification heads and are trained end-to-end with cross-entropy loss using Adam, cosine scheduling, early stopping, and subject-independent 5-fold nested cross-validation.Learning rates are selected from [10^-3 −10^-6], and evaluation reports mean and standard deviation across five outer folds.
  • C. Experimental Setup: Four explainers generate time-channel attributions, which EEG-PRISM maps to spectral and source spaces using independently channel-wise DFTs and an MNE-Python forward model.The explainers are LRP, IG, DeepLIFT, and DeepSHAP; spectral mapping uses fs=200 Hz, while spatial mapping assumes the fsaverage template brain and a three-layer boundary model.

III. RESULTS · A. Simulated Dataset: Comparison with Ground Truth

EEG-PRISM was evaluated on simulated EEG with known spectral and spatial ground truth, enabling direct attribution-mapping accuracy assessment across five foundation models and explainers. Spectral recovery was near-perfect, while spatial attribution localization exceeded chance despite the inverse problem’s high dimensionality and ill-posedness.

  • III. RESULTS: EEG-PRISM was evaluated across several foundation-model configurations on held-out test data, with classification performance reported separately in Appendix I.The method is agnostic to classification and applies to any trained model.
  • A. Simulated Dataset: Comparison with Ground Truth: The simulated dataset encoded known ground-truth spectral and spatial characteristics, while classification was trivial for all five foundation models.This setup enabled quantitative evaluation of attribution recovery and provided confidence that models learned discriminative representations.
  • A. Simulated Dataset: Comparison with Ground Truth: All four explainability methods achieved near-perfect spectral-band accuracy against the ground-truth component, with minor deviations for IG and DeepLift.Accuracy was computed by comparing the maximally attributed frequency band with the band used during signal generation.
  • A. Simulated Dataset: Comparison with Ground Truth: Spatial attribution was mapped onto 5,526 cortical and subcortical sources, reduced from 74,382 forward-model sources to stabilize the ill-posed inverse problem.The evaluation therefore tested attribution recovery in a high-dimensional source space.
  • A. Simulated Dataset: Comparison with Ground Truth: All explainability methods exceeded the 20% chance level for quadrant-level and lobe-level spatial accuracy, with Labram performing best followed by Cbramod.The reported differences between Labram and Cbramod were not statistically significant.
  • A. Simulated Dataset: Comparison with Ground Truth: EEG-PRISM correctly localized attributions to the sources driving the simulated EEG despite source locations being unrelated to task labels, supporting robust post-hoc analysis.An LRP implementation compatible with transformer-based foundation models was unavailable for this evaluation.

B. TUSZ Dataset: Window-Level seizure analysis

EEG-PRISM revealed clinically relevant seizure patterns from window-level EEG foundation-model predictions without directly using spectral or spatial information. It highlighted delta-theta activity and analyzed cortical and subcortical attributions to localize seizure onset zones.

  • Classification and attribution setup: Foundation models achieved seizure-classification accuracies near 88%, comparable to the literature, before EEG-PRISM attribution analysis.The models were trained for window-level seizure-versus-non-seizure classification and did not directly use spectral or spatial information.
  • Spectral analysis: Delta and theta bands consistently received the highest spectral importance, with theta dominant in three of five models.Positive attributions were averaged across five standard frequency bands using correctly classified seizure windows.
  • Spectral analysis: Alpha and beta bands were referenced in 12% and 11% of clinical reports, respectively, and received lower attribution scores, while gamma showed minimal importance.The low gamma importance was attributed to preprocessing that filtered frequencies above 30 Hz.

C. ACE Dataset: Group-Level Autism Analysis · IV. DISCUSSION

EEG-PRISM enables post-hoc, physiologically grounded interpretation of EEG foundation models by mapping time-channel attributions into spectral and spatial domains without modifying or retraining models. In autism analysis and broader evaluation, it identified clinically relevant frequency and regional patterns while demonstrating broad applicability and motivating extensions to joint and nonlinear representations.

  • C. ACE Dataset: Group-Level Autism Analysis: EEG-PRISM analyzes ASD versus TDC using 2-second EEG-window predictions aggregated to the subject level by majority voting.The LaBraM foundation model achieved the highest classification performance, and best-performing models were selected for reliable interpretation.
  • C. ACE Dataset: Group-Level Autism Analysis: Autism spectral analysis averaged positive attributions across five canonical frequency bands and EEG windows, while spatial analysis averaged 5,526 source-level scores across subjects.Analyses were restricted to correctly classified subjects and time windows; spatial scores were interpolated with NiLearn and PyVista.
  • C. ACE Dataset: Group-Level Autism Analysis: ASD classification was primarily associated with increased delta attributions, whereas TDC classification was driven by alpha activity.The passage relates increased delta activity to altered cortical processing and disrupted alpha rhythms to attention and cognitive engagement.
  • IV. DISCUSSION: EEG-PRISM addresses the mismatch between opaque time-channel foundation-model inputs and clinically meaningful spectral and spatial representations.Post-hoc explainers identify predictive input features, but EEG-PRISM extends interpretation to alternative domains without changing the underlying model.
  • IV. DISCUSSION: EEG-PRISM combines signal-processing theory with chain-rule backpropagation and was evaluated across five foundation models and four attribution methods.The work also validated clinical utility in epilepsy through alignment with reported spectral and spatial patterns.
  • IV. DISCUSSION: The invertible DFT enables spectral attribution mapping, yielding near-perfect simulated recovery and consistent identification of delta–theta contributions to seizure detection.The epilepsy finding aligns with established clinical knowledge of epileptiform activity.
  • IV. DISCUSSION: Despite reliance on user-defined forward and inverse models, spatial mapping achieved high simulated accuracy and competitive fully post-hoc seizure-onset-zone localization.Foundation models had no spatial supervision during training, yet EEG-PRISM recovered spatial patterns from their attributions.
  • IV. DISCUSSION: Across datasets, spectral and spatial domains showed more concentrated attribution information than input space, especially in ACE’s frontal regions.The discussion interprets higher mean-normalized ranges as evidence of concentrated predictive features and clinically useful representations.

V. CONCLUSION · PREDICTION TASK

EEG-PRISM is a principled, flexible, fully post hoc, and model-agnostic framework that projects EEG foundation-model attributions into clinically meaningful spectral and spatial domains. Across simulated and real-world datasets, it recovered relevant frequency patterns and spatial regions while supporting competitive seizure characterization and localization and exploratory analyses.

  • V. CONCLUSION: EEG-PRISM projects attribution scores into clinically meaningful spectral and spatial domains.The framework is presented as principled and flexible for interpreting EEG foundation models.
  • V. CONCLUSION: Across simulated and real-world datasets, EEG-PRISM showed strong fidelity in recovering relevant frequency patterns and spatial regions.These recoveries aligned with established neuroscientific and clinical knowledge.
  • V. CONCLUSION: Despite being fully post hoc and model-agnostic, EEG-PRISM achieved competitive performance in seizure characterization and localization.The conclusion frames these results as evidence of the method’s applicability without modifying the underlying model.
  • V. CONCLUSION: EEG-PRISM also showed promise in exploratory analyses.The supplied conclusion identifies exploratory analyses as an additional area of promise without specifying a quantitative result.
  • PREDICTION TASK: Table III reports average classification accuracy for the primary task across five foundation models and the simulated, TUSZ, and ACE datasets.The reported accuracy settings differ by dataset and classification task.
  • PREDICTION TASK: In the simulated dataset, window-level abnormality classification used a 4-way task, while TUSZ used window-level binary seizure classification.The passage states that all models achieved similar performance in these datasets.
  • PREDICTION TASK: In ACE, subject-level ASD-versus-TDC binary classification accuracy was computed after majority voting across window-level model outputs for each EEG recording.The ACE evaluation therefore aggregated window-level predictions at the subject level.
Loading 2608.13676v1…