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

# Physiologically-Grounded Interpretability of Predictions by EEG Foundation Models
Source: [https://arxiv.org/html/2608.13676](https://arxiv.org/html/2608.13676)
Punnisa AmornsirikulArchana VenkataramanThanks:This is a pre\-print\. Manuscript is currently under peer\-reviewThanks:This work was supported in part by NIH R01DC022565 \(PI: Venkataraman\) and NIH R01HD108790 \(PI: Venkataraman\)\.Thanks:D\. M\. Shama is currently with the Department of Electrical and Computer Engineering at Johns Hopkins University, MD, USA 21218 and visiting student researcher at Boston University, MA, USA 02215 \(email: dshama1@jhu\.edu\)Thanks:P\. Amornsirikul is currently with Boston University, MA, USA 02215 \(email: punnisa@bu\.edu\)Thanks:A\. Venkataraman was with Johns Hopkins University and is currently with the Department of Electrical and Computer Engineering at Boston University, MA, USA 02215 \(email: archanav@bu\.edu\)

###### Abstract

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 ofanyfoundation model to alternative and physiologically relevant domainswithout modifying or re\-training 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\.

###### Index Terms:

Foundation models, EEG, Explainability, Domain\-Aware Interpretability, Linear Transformations

## IIntroduction

Scalp electroencephalography \(EEG\) provides a non\-invasive window into brain activity by capturing the rich temporal dynamics of underlying neural processes\. In addition, variations in the structured spatial and spectral activity of EEG signals are closely linked to neurological disorders and pathology\. For example, in the widely\-studied condition of epilepsy, synchronous spiking activity and evolving spectral patterns across frequency bands in the EEG signals provide important insights into seizure dynamics\[[1](https://arxiv.org/html/2608.13676#bib.bib1),[2](https://arxiv.org/html/2608.13676#bib.bib2)\]; they are also used to distinguish seizure subtypes and guide treatment options\[[3](https://arxiv.org/html/2608.13676#bib.bib3),[4](https://arxiv.org/html/2608.13676#bib.bib4)\]\. In the research of psychiatric conditions using EEG\[[5](https://arxiv.org/html/2608.13676#bib.bib5)\], studies have linked alterations in spectral activity, primarily in the lower frequency bands, to disorders such as ADHD and OCD\[[6](https://arxiv.org/html/2608.13676#bib.bib6)\]\. While findings in autism remain mixed, there is some evidence suggesting differences in higher\-frequency \(e\.g\., gamma\) activity\[[7](https://arxiv.org/html/2608.13676#bib.bib7)\]\.

In the spatial domain, EEG signals are modeled as the result of electrical activity generated by distributed neural sources which propagate through head tissues to the scalp according to a well\-established forward model\[[8](https://arxiv.org/html/2608.13676#bib.bib8)\]\. Particularly with high density recordings, there is increasing interest in leveraging these spatial relationships to “invert" the forward model and map the EEG activity to cortical regions in the brain\. For example, scalp EEG can offer approximate localization of seizure onset regions and serve as a noninvasive precursor to invasive evaluation of epileptogenic zones\[[9](https://arxiv.org/html/2608.13676#bib.bib9)\]\. Source\-level interpretations are also increasingly leveraged for psychiatric conditions\[[10](https://arxiv.org/html/2608.13676#bib.bib10)\]\.

Despite its clinical value, EEG is complex, high\-dimensional, and often contaminated by artifacts\. These factors make manual interpretation challenging, particularly for continuous and multichannel recordings\. The rise of artificial intelligence has led to computational methods for EEG analysis, with increasing emphasis on learning representations directly from EEG data that generalize across datasets and tasks\[[11](https://arxiv.org/html/2608.13676#bib.bib11)\]\. While these models can achieve high prediction accuracies, their clinical utility is limited by their lack of explainability, as shown in Fig\.[1](https://arxiv.org/html/2608.13676#S1.F1)\. In particular, these models operate on EEG signals as input, so post\-hoc explainers, such as GradCam\[[12](https://arxiv.org/html/2608.13676#bib.bib12)\]and SHAP\[[13](https://arxiv.org/html/2608.13676#bib.bib13)\], provide temporal and channel\-level attribution scores\. This information does not align with how EEG is understood in practice, where diagnostic insights are more naturally expressed in the spectral and spatial domains\. Thus, there is a growing need for trustworthy approaches whose interpretations are not only accurate, but also grounded in physiologically meaningful representations that can integrate seamlessly into clinical workflows\.

![Refer to caption](https://arxiv.org/html/2608.13676v1/problem.jpg)Fig\. 1:EEG foundation models and their explainability using post\-hoc tools operate in time\-channel space, while clinically meaningful patterns lie in spectral and source domains that govern EEG generation\.### I\-AEEG Foundation Models as Black\-Box Computational Tools

EEG foundation models are increasingly replacing traditional feature\-engineering pipelines due to their improved performance and ease of deployment\. Traditional approaches are often pipelined to include preprocessing, handcrafted feature extraction, and a prediction module \(e\.g\., logistic regression or support vector machines\)\. The handcrafted features \(e\.g\., Fourier transforms, wavelet decompositions, and spectral power\) are motivated by neuroscientific insights, but they require extensive experimentation to identify representations that best capture the phenomenon of interest\[[14](https://arxiv.org/html/2608.13676#bib.bib14)\]\. In contrast, deep neural networks learn hierarchical representations directly from the time–channel EEG inputs and reduce the need for manual feature design\[[15](https://arxiv.org/html/2608.13676#bib.bib15)\]\. With sufficient training data, these networks can achieve greater predictive performance than traditional approaches, and they often demonstrate improved generalization across datasets\[[16](https://arxiv.org/html/2608.13676#bib.bib16)\]\.

Foundation models represent the next advancement in AI for EEG analysis\. They build on early deep neural networks by \(1\) pretraining the model on large\-scale EEG datasets that comprise thousands of hours of recordings; and \(2\) using self\-supervised learning to provide flexibility across downstream prediction tasks\. This setup is in contrast to earlier deep neural networks, including convolutional architectures such as EEGNet\[[17](https://arxiv.org/html/2608.13676#bib.bib17)\]and hybrid convolution–transformer models such as EEGConformer\[[18](https://arxiv.org/html/2608.13676#bib.bib18)\], which require task\-specific training from scratch\. Current EEG foundation models adopt a range of deep neural network architectures for large\-scale self\-supervised representation learning\. All models learn the self\-supervised representations directly from the EEG signals \(i\.e\., time\-channel inputs\)\. Early models, such as BENDR\[[19](https://arxiv.org/html/2608.13676#bib.bib19)\], adapt a contrastive prediction framework from speech analysis to EEG, while more recent models like BIOT\[[20](https://arxiv.org/html/2608.13676#bib.bib20)\]employ masked training to learn temporal dependencies from a large corpora\. Other works focus on improving input representations, including Labram\[[21](https://arxiv.org/html/2608.13676#bib.bib21)\], which introduces discrete neural tokenization, and FoME\[[22](https://arxiv.org/html/2608.13676#bib.bib22)\], which incorporates time–frequency fusion to better capture spectral structure\. Models such as Cbramod\[[23](https://arxiv.org/html/2608.13676#bib.bib23)\]explicitly separate spatial and temporal attention mechanisms, whereas LUNA\[[24](https://arxiv.org/html/2608.13676#bib.bib24)\]addresses data heterogeneity by learning topology\-agnostic latent representations\.

While powerful, foundation models operate as black\-box units and learn representations directly from the time\-channel input data\. This strategy enables rapid adaptation and strong performance across datasets, but their decision mechanisms remain difficult to interpret\. Moreover, current explainable AI techniques attempt to quantify input\-output relationships and cannot provide explanations in the clinically meaningful spectral and spatial domains\. On the other hand, incorporating domain\-relevant information into the foundation model is computationally expensive, requiring architectural changes and retraining\. Taken together, there is a need for an explainability tool that directly extract physiologically\-grounded explanations with minimal overhead as a crucial stepping stone towards clinical translation\.

### I\-BExplainability in Deep Learning Pipelines

Current explainable AI methods do not aim for full transparency of black\-box models, but rather to address the question ofwhy a particular prediction is made\. Given the difficulty of modifying black\-box deep neural networks, many widely used methods are post hoc, operating after model training and inference\. These methods are typically model\-agnostic and assign “attribution scores" to quantify the contribution of each input feature to the model prediction\. Notably for EEG foundation models, the attribution scores are provided on the time\-channel inputs, which may be difficult to interpret in the context of a neurological disorder\[[25](https://arxiv.org/html/2608.13676#bib.bib25),[26](https://arxiv.org/html/2608.13676#bib.bib26)\]\.

At a high level, post\-hoc explainers use specific rules to backpropagate information through the neural network in order to compute the attribution scores\. Popular methods include:

- •Layer\-wise Relevance Propagation \(LRP\)redistributes the model prediction backward through the network according to conservation principles involving different propagation rules, such as theϵ\\epsilon\-rule andγ\\gamma\-rule for neural networks, as well as customized rules for transformers\[[27](https://arxiv.org/html/2608.13676#bib.bib27),[28](https://arxiv.org/html/2608.13676#bib.bib28),[29](https://arxiv.org/html/2608.13676#bib.bib29),[30](https://arxiv.org/html/2608.13676#bib.bib30),[31](https://arxiv.org/html/2608.13676#bib.bib31)\]\.
- •Integrated Gradientsconstructs attribution scores by integrating the gradients of the model output with respect to the input along a path from a baseline to the observed sample\. This strategy ensures that the attributions are highly sensitive and invariant to implementation across equivalent models\[[32](https://arxiv.org/html/2608.13676#bib.bib32),[33](https://arxiv.org/html/2608.13676#bib.bib33),[34](https://arxiv.org/html/2608.13676#bib.bib34)\]\.
- •DeepLIFTcomputes the attribution scores by comparing neuron activations to a reference baseline and propagating these differences through the network\. This approach allows for non\-zero attributions even in regions where gradients are saturated, providing more robust attribution scores\[[35](https://arxiv.org/html/2608.13676#bib.bib35),[36](https://arxiv.org/html/2608.13676#bib.bib36),[37](https://arxiv.org/html/2608.13676#bib.bib37)\]
- •SHAPassigns input\-level scores that correspond to theoretically\-grounded Shapley values\. DeepSHAP extends the original framework to deep networks by combining SHAP’s attribution principle with DeepLIFT\-style propagation\. This combination enables the scalable approximation of Shapley\-consistent attributions in complex models\[[38](https://arxiv.org/html/2608.13676#bib.bib38),[39](https://arxiv.org/html/2608.13676#bib.bib39),[40](https://arxiv.org/html/2608.13676#bib.bib40),[41](https://arxiv.org/html/2608.13676#bib.bib41),[13](https://arxiv.org/html/2608.13676#bib.bib13)\]\.

Given that EEG foundation models operate directly on time–channel inputs, attribution scores obtained using any of these post\-hoc methods are expressed in the time–channel domain\. However, such representations are not always interpretable or actionable by a clinician\. For example, temporal attributions at the millisecond scale may not align with clinically relevant patterns, which often span longer durations\. Similarly, channel\-wise attributions can be misleading due to strong correlations between electrodes\. As previously discussed, clinical phenomena in EEG are more often associated with spectral \(frequency\-domain\) and spatial \(source\-domain\) patterns\.

### I\-COur Contributions

In this paper, we introduce EEG\-PRISM forPhysiologically\-gRoundedInterpretability of PredictionSby EEG foundationModels\. Our work includes several novel contributions for EEG analysis:

- •Amathematically grounded frameworkto map EEG foundation model attributions into clinically meaningful frequency and source domains using DFT and inverse modeling\. Our theoretical contributions include a derived error bound and guaranteed compatibility across multiple explainers via backpropagation\.
- •Amodel\-agnostic and post\-hoc approachthat does not require any modification to the existing foundation model; EEG\-PRISM also has minimal computational overhead\.
- •Extensive simulation studieswith known ground truth in the spectral and source domains demonstrate the attribution fidelity of EEG\-PRISM across five different EEG foundation models\.
- •Real\-world biomedical validationin seizure localization and autism biomarker discovery\. EEG\-PRISM enables both subject\-specific event detection and group\-level neural characterization\.

Taken together, we present a valuable tool that aligns the trajectory of AI\-based EEG analysis with clinical and research practice\.

![Refer to caption](https://arxiv.org/html/2608.13676v1/prism.jpg)Fig\. 2:Overall pipeline of EEG\-PRISM for explainability in clinically relevant domains beyond the input time\-channel space\.Top:End\-to\-end experimental setup, in which EEG\-PRISM is applied to foundation models post\-hoc to obtain attributions in the spectral and spatial domains\.Bottom left:Transforming attributions to new domains within EEG\-PRISM\.Bottom right:Potential downstream applications\.

## IIMethods

EEG\-PRISM leverages the principle ofattribution propagationto map explanations derived from foundation models that operate on the time–channel EEG input space onto clinically meaningful subspaces\. Mathematically, letX∈ℝC×TX\\in\\mathbb\{R\}^\{C\\times T\}denote the input EEG, whereCCis the number of channels andTTthe number of time points\. The AI foundation modelG⁡\(⋅\)G\(\\cdot\)usesXXfor a prediction task \(e\.g\., case/control classification\), leading to post\-hoc attributions𝒜∈ℝC×T\\mathcal\{A\}\\in\\mathbb\{R\}^\{C\\times T\}in the input space\. Our goal is to transform𝒜\\mathcal\{A\}into \(1\) spectral attributions𝒜F\\mathcal\{A\}^\{F\}that capture predictive frequency components, and \(2\) spatial attributions𝒜S\\mathcal\{A\}^\{S\}that capture importance across source locations\. Crucially, the transformation should not require modification or retraining ofG⁡\(⋅\)G\(\\cdot\)\. Noting that the spectral domain can be reached from the input space via an invertible discrete Fourier transform, while the spatial domain arises by inverting the EEG forward model, the full pipeline for EEG\-PRISM is shown in Fig\.[2](https://arxiv.org/html/2608.13676#S1.F2)\.

### II\-AMapping Attribution Scores to Clinically Relevant Domains

Consider a generic linear transform between the input and desired subspace\. Formally, let𝐱∈ℝN\\mathbf\{x\}\\in\\mathbb\{R\}^\{N\}denote a vector in either the row space \(N = T\) or column space \(N = C\) of the EEG inputX∈ℝC×TX\\in\\mathbb\{R\}^\{C\\times T\}\. Let𝐳∈ℝM\\mathbf\{z\}\\in\\mathbb\{R\}^\{M\}withM≥NM\\geq Nrepresent the corresponding vector in the target subspace, such that𝐱=W​𝐳\\mathbf\{x\}=W\\mathbf\{z\}, whereW∈ℝN×MW\\in\\mathbb\{R\}^\{N\\times M\}is a well\-defined linear transformation\. EEG\-PRISM leverages the following proposition to directly map the attribution scores:

###### Proposition 1

The attribution scores in thezz\-space, denoted by𝒜z\\mathcal\{A\}^\{z\}, can be expressed as signal times a weighted linear combination of the original attribution scores\. Let𝒜x\\mathcal\{A\}^\{x\}denote the row or column of𝒜\\mathcal\{A\}corresponding to the input vector𝐱\\mathbf\{x\}\. Then we have:

𝒜iz=𝐳i​∑j=1NWj​i​𝒜jx𝐱j,\\mathcal\{A\}^\{z\}\_\{i\}=\\mathbf\{z\}\_\{i\}\\sum\_\{j=1\}^\{N\}W\_\{ji\}\\frac\{\\mathcal\{A\}^\{x\}\_\{j\}\}\{\\mathbf\{x\}\_\{j\}\},\(1\)where the weightsWj​iW\_\{ji\}correspond to the linear transformation\.

###### Proof:

Treating the subspace transformation as a linear change of variables, we define the composite functionG~​\(𝐳\):=G⁡\(W​𝐳\)=G⁡\(𝐱\)=y\\tilde\{G\}\(\\mathbf\{z\}\):=G\(W\\mathbf\{z\}\)=G\(\\mathbf\{x\}\)=y\. SinceWWis fixed and linear, all attribution rules of post\-hoc explainers apply through the standard composition\. We outline the proof for three representative classes of explainers\.

##### Layer\-wise Relevance Propagation \(LRP\)

The proof of Proposition 1 for LRP is detailed in\[[42](https://arxiv.org/html/2608.13676#bib.bib42)\]and restated here for completeness\. Using standard relevance propagation for linear layers and the conservation principle, we obtain the relationship:

𝒜iz=∑j=1NWj​i​𝐳i∑kWj​k​𝐳k​𝒜jx\.\\mathcal\{A\}^\{z\}\_\{i\}=\\sum\_\{j=1\}^\{N\}\\frac\{W\_\{ji\}\\mathbf\{z\}\_\{i\}\}\{\\sum\_\{k\}W\_\{jk\}\\mathbf\{z\}\_\{k\}\}\\mathcal\{A\}^\{x\}\_\{j\}\.\(2\)Note that Eq\. \([2](https://arxiv.org/html/2608.13676#S2.E2)\) is equivalent to Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) as∑kWj​k​𝐳k=𝐱j\\sum\_\{k\}W\_\{jk\}\\mathbf\{z\}\_\{k\}=\\mathbf\{x\}\_\{j\}\.

##### Gradient\-Based Explainers

This proof hinges on the chain rule\. Let us consider the simplest method of Gradient times Input \(G×\\timesI\)\. The chain rule over the composite functionG~​\(𝐳\)\\tilde\{G\}\(\\mathbf\{z\}\)yields

∂G~∂𝐳i=∑j=1N∂G∂𝐱j⋅∂𝐱j∂𝐳i=∑j=1N∂G∂𝐱j⋅Wj​i\.\\frac\{\\partial\\tilde\{G\}\}\{\\partial\\mathbf\{z\}\_\{i\}\}=\\sum\_\{j=1\}^\{N\}\\frac\{\\partial G\}\{\\partial\\mathbf\{x\}\_\{j\}\}\\cdot\\frac\{\\partial\\mathbf\{x\}\_\{j\}\}\{\\partial\\mathbf\{z\}\_\{i\}\}=\\sum\_\{j=1\}^\{N\}\\frac\{\\partial G\}\{\\partial\\mathbf\{x\}\_\{j\}\}\\cdot W\_\{ji\}\.\(3\)
Thus, the attribution propagates linearly throughWWas follows:

Aiz=𝐳i​∂G~​\(𝐳\)∂𝐳i=𝐳i​∑j∂G⁡\(𝐱\)∂𝐱j​Wj​i=𝐳i​∑jWj​i​𝒜jx𝐱jA^\{z\}\_\{i\}=\\mathbf\{z\}\_\{i\}\\frac\{\\partial\\tilde\{G\}\(\\mathbf\{z\}\)\}\{\\partial\\mathbf\{z\}\_\{i\}\}=\\mathbf\{z\}\_\{i\}\\sum\_\{j\}\\frac\{\\partial\{G\}\(\\mathbf\{x\}\)\}\{\\partial\\mathbf\{x\}\_\{j\}\}W\_\{ji\}=\\mathbf\{z\}\_\{i\}\\sum\_\{j\}W\_\{ji\}\\frac\{\\mathcal\{A\}^\{x\}\_\{j\}\}\{\\mathbf\{x\}\_\{j\}\}\(4\)
Consider Integrated Gradients \(IG\), which reduces to a path\-integral accumulation of input gradients that satisfies crucial axioms of completeness and implementation invariance\[[32](https://arxiv.org/html/2608.13676#bib.bib32)\]\. By combining the chain rule in Eq\. \([3](https://arxiv.org/html/2608.13676#S2.E3)\) and the linearity property of integration givenx=W​zx=Wz, we obtain the following attributions for IG:

Aiz\\displaystyle A^\{z\}\_\{i\}=𝐳i​∫α∂G~​\(α​𝐳\)∂𝐳i​𝑑α=𝐳i​∫α∑j∂G⁡\(α​𝐱\)∂𝐱j​Wj​i​𝑑α\\displaystyle=\\mathbf\{z\}\_\{i\}\\int\_\{\\alpha\}\\frac\{\\partial\\tilde\{G\}\(\\alpha\\mathbf\{z\}\)\}\{\\partial\\mathbf\{z\}\_\{i\}\}d\\alpha=\\mathbf\{z\}\_\{i\}\\int\_\{\\alpha\}\\textstyle\\sum\_\{j\}\\frac\{\\partial\{G\}\(\\alpha\\mathbf\{x\}\)\}\{\\partial\\mathbf\{x\}\_\{j\}\}W\_\{ji\}\\,d\\alpha=𝐳i​∑j=1NWj​i​𝒜jx𝐱j\\displaystyle=\\mathbf\{z\}\_\{i\}\\sum\_\{j=1\}^\{N\}W\_\{ji\}\\frac\{\\mathcal\{A\}^\{x\}\_\{j\}\}\{\\mathbf\{x\}\_\{j\}\}\(5\)Note that the final step is a direct result of the definition of Integrated Gradients in thexx\-space,𝒜jx=𝐱j​∫α∂G⁡\(α​𝐱\)∂𝐱j​𝑑α\\mathcal\{A\}^\{x\}\_\{j\}=\\mathbf\{x\}\_\{j\}\\int\_\{\\alpha\}\\frac\{\\partial\{G\}\(\\alpha\\mathbf\{x\}\)\}\{\\partial\\mathbf\{x\}\_\{j\}\}d\\alpha\.

##### DeepLIFT and DeepSHAP

The proof for this class of explainers uses the backpropagation and chain rules of discrete differentialsmm\. This formulation leads to linear redistribution under fixed affine transformations\. Specifically, the chain rule states

mΔ𝐳i​Δy=∑jmΔ𝐳i​Δ𝐱j​mΔ𝐱j​Δy,m\_\{\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\Delta\_\{y\}\}=\\sum\_\{j\}m\_\{\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\}m\_\{\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\\Delta\_\{y\}\},\(6\)whereΔ\(⋅\)\\Delta\_\{\(\\cdot\)\}is the difference from a fixed reference\. For the linear relation𝐱=W​𝐳\\mathbf\{x\}=W\\mathbf\{z\}, DeepLIFT and DeepSHAP assignmΔ𝐳i​Δ𝐱j=Wj​im\_\{\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\}=W\_\{ji\}\. Further, using the equalitymΔ𝐱j​Δy=𝒜jxΔ𝐱jm\_\{\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\\Delta\_\{y\}\}=\\frac\{\\mathcal\{A\}^\{x\}\_\{j\}\}\{\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\}we have

𝒜iz=Δ𝐳i⋅mΔ𝐳i​Δy=Δ𝐳i⋅∑jWj​i​𝒜j𝐱Δ𝐱j\.\\mathcal\{A\}^\{z\}\_\{i\}=\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\cdot m\_\{\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\Delta\_\{y\}\}=\\Delta\_\{\\mathbf\{z\}\_\{i\}\}\\cdot\\sum\_\{j\}W\_\{ji\}\\frac\{\\mathcal\{A\}^\{\\mathbf\{x\}\}\_\{j\}\}\{\\Delta\_\{\\mathbf\{x\}\_\{j\}\}\}\.\(7\)By setting the reference to zero, i\.e\.,Δ𝐳i=𝐳i\\Delta\_\{\\mathbf\{z\}\_\{i\}\}=\\mathbf\{z\}\_\{i\}andΔ𝐱i=𝐱i\\Delta\_\{\\mathbf\{x\}\_\{i\}\}=\\mathbf\{x\}\_\{i\}, Eq\. \([7](https://arxiv.org/html/2608.13676#S2.E7)\) becomes equivalent to Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) in the proposition\.

For DeepSHAP, which builds on DeepLIFT, the proof follows the same structure with empirical average over multiple references\.□\\square∎

![Refer to caption](https://arxiv.org/html/2608.13676v1/simulation.jpg)Fig\. 3:Simulated EEG data generation pipeline using SEREEGA\. \(1\) Initialize the head model and select a source location, \(2\) Assign class\-associated frequency signal, \(3\) Generate EEG from source activity using the leadfield matrix, and \(4\) Preprocess the EEG using standard methods\.Taken together, Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) is theoretically grounded and preserves the essential properties of the base explainers \(e\.g\., faithfulness, conservation, and stability\) without modifyingG⁡\(⋅\)G\(\\cdot\)\. EEG\-PRISM leverages this result to map time\-channel input attribution scores to the spectral and spatial domains as described below\.

#### II\-A1Spectral Domain

For each EEG channel, the spectral representation𝐳c1:T\\mathbf\{z\}^\{c\}\_\{1:T\}of the EEG signal𝐱c1:T\\mathbf\{x\}^\{c\}\_\{1:T\}is obtained via the Discrete Fourier transform \(DFT\), which is a predefined linear and invertible operation\. The caveat is that a DFT produces complex\-valued coefficients, which would, in turn, lead to complex\-valued attribution scores due to the multiplication in Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\)\. As seen in\[[42](https://arxiv.org/html/2608.13676#bib.bib42)\]we decompose the DTF coefficients𝐳c1:T\\mathbf\{z\}^\{c\}\_\{1:T\}into its real and imaginary components\. The inverse DFT to reconstruct the EEG data𝐱c1:T\\mathbf\{x\}^\{c\}\_\{1:T\}can be written

𝐱jc=∑i=1TRe​\(𝐳ic\)​cos​\(2​π​j​iT\)−Im​\(𝐳ic\)​sin​\(2​π​j​iT\)\.\\mathbf\{x\}^\{c\}\_\{j\}=\\sum\_\{i=1\}^\{T\}\\text\{Re\}\(\\mathbf\{z\}^\{c\}\_\{i\}\)\\text\{cos\}\\left\(\\frac\{2\\pi ji\}\{T\}\\right\)\-\\text\{Im\}\(\\mathbf\{z\}^\{c\}\_\{i\}\)\\text\{sin\}\\left\(\\frac\{2\\pi ji\}\{T\}\\right\)\.\(8\)whereRe​\(⋅\)\\text\{Re\}\(\\cdot\)andIm​\(⋅\)\\text\{Im\}\(\\cdot\)denote the real and imaginary components, respectively\. From here, we create a new vector representation by concatenating the real and imaginary parts:\[Re\(zc\)\|Im\(zc\)\]\[\\text\{ Re\}\(z^\{c\}\)\\text\{ \}\|\\text\{ Im\}\(z^\{c\}\)\\text\{ \}\]\. The linear weightWi​jW\_\{ij\}to map the attributions from the time domain to the spectral domain is obtained from Eq\. \([8](https://arxiv.org/html/2608.13676#S2.E8)\) ascos⁡\(2​π​j​i/T\)\\cos\\\!\\left\(\{2\\pi ji/T\}\\right\)for the real part and−sin⁡\(2​π​j​i/T\)\-\\sin\\\!\\left\(\{2\\pi ji/T\}\\right\)for the imaginary part of the vector\.

Substituting these expressions into Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) and adding the attribution scores of the real and imaginary components gives us the spectral attributions per channel as follows:

𝒜iF=1C​∑c=1C∑j=1T\[Re​\(𝐳ic\)​cos​\(2​π​j​iT\)−Im​\(𝐳ic\)​sin​\(2​π​j​iT\)\]​𝒜jcxjc\\mathcal\{A\}^\{F\}\_\{i\}=\\frac\{1\}\{C\}\\sum\_\{c=1\}^\{C\}\\sum\_\{j=1\}^\{T\}\\left\[\\text\{Re\}\(\\mathbf\{z\}^\{c\}\_\{i\}\)\\text\{cos\}\\left\(\\frac\{2\\pi ji\}\{T\}\\right\)\-\\text\{Im\}\(\\mathbf\{z\}^\{c\}\_\{i\}\)\\text\{sin\}\\left\(\\frac\{2\\pi ji\}\{T\}\\right\)\\right\]\\frac\{\\mathcal\{A\}^\{c\}\_\{j\}\}\{x^\{c\}\_\{j\}\}\(9\)
Finally, averaging across all electrode channels provides the overall spectral\-domain attribution𝒜F\\mathcal\{A\}^\{F\}for the EEG data\.

#### II\-A2Spatial Domain

The relationship between the multichannel EEG data at time pointtt, i\.e\.,𝐱1:Ct\\mathbf\{x\}^\{1:C\}\_\{t\}and the corresponding spatial source domain𝐳tS\\mathbf\{z\}^\{S\}\_\{t\}is modeled via the EEG forward model𝐱1:Ct=W𝐳St\\mathbf\{x\}^\{1:C\}\_\{t\}=W\\mathbf\{z\}^\{S\}\_\{t\}\. Here,𝐳tS∈ℝM\\mathbf\{z\}^\{S\}\_\{t\}\\in\\mathbb\{R\}^\{M\}denotes the instantaneous source\-level \(e\.g\., cortical\) activity acrossMMsources, andW∈ℝC×MW\\in\\mathbb\{R\}^\{C\\times M\}is the forward matrix derived from biophysically informed head conductivity models that assume a fixed normal dipole orientation\. Since the EEG forward problem provides a well\-defined linear mapping, the transformation weightsWj​iW\_\{ji\}in Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) can be directly obtained from the in\-built matrix\. A straightforward substitution of these weights along withztSz^\{S\}\_\{t\}yields instantaneous attribution scores, which are then averaged across time to obtain the source domain attribution,𝒜S\\mathcal\{A\}^\{S\}\.

The primary consideration in applying Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\) is that of estimated the source\-level activity lies in obtaining the source activity𝐳tS\\mathbf\{z\}^\{S\}\_\{t\}required for the computation\. Estimating𝐳tS\\mathbf\{z\}^\{S\}\_\{t\}corresponds to the EEG inverse problem, for which a range of well\-established solvers exist\. Since the number of cortical sources is often greater than the number of EEG channels, i\.e\.,M≫CM\\gg C, the problem is inherently ill\-posed and requires appropriate regularization to invert\.

Importantly, EEG\-PRISM does not amplify or introduce errors into the inverse mapping\. Rather, EEG\-PRISM operates post\-hoc and propagates attribution scores through a linear mapping\. As a result, any error in the estimated source𝐳S\\mathbf\{z\}^\{S\}is transferred linearly, with no additional distortion beyond a constant scaling factor determined by the transformation\. Formally, if the true source𝐳true\\mathbf\{z\}^\{\\text\{true\}\}differs from the estimated𝐳S\\mathbf\{z\}^\{S\}, then the induced error in the mapped attribution remains proportional to this discrepancy as follows:

‖𝒜t​r​u​e−𝒜S‖≤‖zt​r​u​e−zS‖⋅‖∑j=1Nwj​i​𝒜jxj‖⏟constant\\left\\\|\\mathcal\{A\}^\{true\}\-\\mathcal\{A\}^\{S\}\\right\\\|\\leq\\left\\\|z^\{true\}\-z^\{S\}\\right\\\|\\cdot\\underbrace\{\\left\\\|\\sum\_\{j=1\}^\{N\}w\_\{ji\}\\frac\{\\mathcal\{A\}^\{j\}\}\{x^\{j\}\}\\right\\\|\}\_\{\\text\{constant\}\}\(10\)Thus, EEG\-PRISM can seamlessly project attributions to the spatial domain, while preserving linear fidelity and not introducing any distortion beyond what is inherent in the inverse solver\.

### II\-BDatasets

Our experiments include three datasets: a simulated dataset for quantitative evaluation of EEG\-PRISM against a known ground truth, and two real\-world datasets used to demonstrate clinical potential\.

#### II\-B1Simulated Dataset

We generate EEG using SEREEGA\[[43](https://arxiv.org/html/2608.13676#bib.bib43)\]with its built\-in source\-to\-scalp forward modeling framework, as shown in Fig\.[3](https://arxiv.org/html/2608.13676#S2.F3)\. A 32\-channel Biosemi electrode montage is employed, along with the ICBM New York head model for EEG forward modeling with 74,382 sources\[[44](https://arxiv.org/html/2608.13676#bib.bib44)\]\. We construct “classes" with labels that are explicitly tied to the underlying source\-level signal characteristics\. This strategy enables a controlled evaluation of EEG\-PRISM across different settings\. The dataset consists of four classes, each defined by a distinct frequency band in the source signal domain: delta/theta \(1–8 Hz\), alpha \(8–12 Hz\), beta \(13–30 Hz\), and gamma \(30–100 Hz\)\. For each sample, a source is randomly selected from the brain, and its activity is assigned oscillatory components within the frequency range associated with its class\. The resulting signals are projected to the scalp to form the time–channel EEG inputs\. A total of 100 samples per class \(400 total\) are generated\. Thus, classification depends on learning frequency\-specific representations internally within the foundation model from the time\-channel input\.

#### II\-B2TUSZ Dataset

We selected 124 subjects \(58 M, 66 F\) with focal seizure onsets from the publicly accessible Temple University Hospital Seizure \(TUSZ\) corpus\[[45](https://arxiv.org/html/2608.13676#bib.bib45)\]\. Patients ranged in age from 19 to 91 years \(55\.2±\\pm16\.6 years\)\. The TUSZ corpus contains several 19\-channel average\-referenced EEG recordings per epilepsy patient, along with expert annotations of seizure intervals\. We randomly crop 30 seconds of non seizure data and up to 30 seconds of seizure data from every EEG recording of every patient; the cropped data is segmented into non\-overlapping 2s windows, resulting in 39,123 non seizure and 18,603 seizure windows in our final TUSZ dataset\.

The EEG foundation models are adapted for binary classification of seizure versus non\-seizure windows\. We use the unstructured clinical notes to extract the seizure onset channels and relevant signal frequency bands for each patient\. The patient\-level seizure onsets are organized as: 31 left frontal, 14 right frontal, 26 left posterior, 45 right posterior, and 8 central\. While the EEG foundation models do not incorporate the onset location, this spatial information is clinically relevant\. Therefore, we quantify the alignment between the spatial attributions of EEG\-PRISM and the clinician\-determined onsets\.

#### II\-B3ACE Dataset

We use resting\-state EEG data from the Autism Center for Excellence \(ACE\) project \(NDAR study \#2021\)\[[46](https://arxiv.org/html/2608.13676#bib.bib46)\]; the dataset includes participants with autism spectrum disorder \(ASD\) and typically developing controls \(TDC\)\. Of the 339 enrolled participants, N=177 \(87 ASD, 90 TDC\) met our inclusion criteria, which required artifact\-free EEG, valid diagnostic and behavioral assessments, and no familial relatedness\. Data were acquired using a 128\-channel EGI Net Amps 300 system with HydroCel nets\. Recordings consisted of alternating task blocks, from which eyes\-open resting\-state EEG segments were extracted for analysis\.

The EEG foundations models are adapted for subject\-level ASD versus TDC classification using all the available time\-channel EEG per participant\. We evaluate the spectral and spatial attribution maps provided by EEG\-PRISM against findings in the ASD literature\.

#### Preprocessing

We apply a uniform preprocessing pipeline across all datasets\. The EEG signals are downsampled to 200 Hz and segmented into 5\-second windows for the simulated dataset, and 2\-second windows for real\-word datasets\. The 5\-second windows in the simulated dataset is selected to match the default configurations of the EEG foundation models\. The smaller window size in real\-word datasets were chosen to improve the seizure detection resolution and to match the experimental protocol of ACE dataset\. We applied a band\-pass filter between 0\.1\-100 Hz for the simulated and ACE datasets and between 0\.1\-30 Hz for the TUSZ dataset\. We also apply a 60 Hz notch filter to remove power line noise\. Finally, we clip the signals at two standard deviations from mean to remove high intensity artifacts\. The time\-channel signals are normalized to have zero mean and unit variance and input to the EEG foundation models\.

### II\-CExperimental Setup

The experimental setup comprises of three main steps: \(i\) Training the EEG foundation models \(ii\) Computing input attribution scores and \(iii\) Mapping the scores to spectral and spatial domains\. Our code will be available in our GitHub repository111https://github\.com/deeksha\-ms/EEG\-PRISMupon acceptance\.

Foundation Models:We apply EEG\-PRISM to models encompassing five deep network architectures\. The two task\-specific models, EEGNet\[[17](https://arxiv.org/html/2608.13676#bib.bib17)\]and AttentionBaseNet \(ABNet\)\[[47](https://arxiv.org/html/2608.13676#bib.bib47)\], are trained from scratch\. Three larger generic models, LaBraM\[[21](https://arxiv.org/html/2608.13676#bib.bib21)\], CBramod\[[23](https://arxiv.org/html/2608.13676#bib.bib23)\], and LUNA\[[24](https://arxiv.org/html/2608.13676#bib.bib24)\], are initialized from released weights and subsequently fine\-tuned on each dataset\. All models are implemented using the Braindecode package\[[48](https://arxiv.org/html/2608.13676#bib.bib48)\]with default configurations as provided in their original releases, ensuring reproducibility across experiments\.

Training Setup:For each dataset, we augment the backbone foundation model with a task\-specific classification head and train the entire network in an end\-to\-end fashion using a cross\-entropy loss\.

We use the Adam optimizer\[[49](https://arxiv.org/html/2608.13676#bib.bib49)\]for training with with cosine learning rate scheduling\. Training is capped at a maximum of 50 epochs, with early stopping based on the validation set performance to prevent overfitting\. We follow a subject\-independent 5\-fold nested cross\-validation scheme\. Within each outer fold, an inner train–validation split is used to select the optimal learning rate within the \[10−3−10−610^\{\-3\}\-10^\{\-6\}\] range\. The selected model is then retrained using both the training and validation sets and is evaluated on the held\-out test fold\. Performance is reported as the mean and standard deviation of evaluation metrics across the five outer folds\. All experiments are implemented in PyTorch \(v2\.11\) and executed on a single A100 GPU\.

Attribution Scores and Mapping with EEG\-PRISM:We select the best\-performing model in every test\-fold and compute the input\-level attributions scores using LRP, IG, DeepLIFT, and DeepSHAP in the time–channel domain\. LRP is implemented using Zennit package while the others are implemented using the Captum Python package\[[50](https://arxiv.org/html/2608.13676#bib.bib50)\]\. IG and DeepLIFT uses a zero reference, while DeepSHAP uses random samples from training set as reference\.

We use EEG\-PRISM to transform each set of attribution scores into both spectral and source\-space attributions\. For spectral mapping, we implement the DFT and its inverse from scratch in PyTorch in order to extract its coefficients in Eq\. \([1](https://arxiv.org/html/2608.13676#S2.E1)\)\. The DFT is applied independently to each channel using a sampling frequency offsf\_\{s\}=200 Hz\. For spatial mapping, we use MNE\-Python\[[51](https://arxiv.org/html/2608.13676#bib.bib51)\]to compute the forward model coefficients while assuming the standardfsaveragetemplate brain and a three\-layer boundary element method \(BEM\) head model\. The inverse solution is computed using sLORETA with the default regularization settings in MNE\[[52](https://arxiv.org/html/2608.13676#bib.bib52),[53](https://arxiv.org/html/2608.13676#bib.bib53)\]\.

TABLE I:Accuracy of EEG\-PRISM derived attributions vs ground\-truth features in simulated dataset\.Left:Comparison between the top spectral band and ground truth band\.Right:Comparison between the center of mass location and ground truth source\.Foundation ModelSpectral DomainSpatial/Source DomainLRP†\\daggerIGDeepLIFTDeepSHAPLRP†\\daggerIGDeepLIFTDeepSHAPEEGNet1\.000±\\pm0\.0000\.942±\\pm0\.0370\.940±\\pm0\.0391\.000±\\pm0\.0000\.613±\\pm0\.0430\.655±\\pm0\.0490\.660±\\pm0\.0520\.600±\\pm0\.080ABNet1\.000±\\pm0\.0000\.988±\\pm0\.0080\.978±\\pm0\.0151\.000±\\pm0\.0000\.655±\\pm0\.0790\.658±\\pm0\.0880\.655±\\pm0\.0890\.595±\\pm0\.088Labram†–0\.960±\\pm0\.0220\.972±\\pm0\.0221\.000±\\pm0\.000–0\.692±\\pm0\.0540\.680±\\pm0\.0630\.638±\\pm0\.072Cbramod†–0\.993±\\pm0\.0100\.995±\\pm0\.0061\.000±\\pm0\.000–0\.670±\\pm0\.0570\.638±\\pm0\.0790\.628±\\pm0\.090LUNA†–0\.945±\\pm0\.0200\.980±\\pm0\.0130\.998±\\pm0\.005–0\.648±\\pm0\.0930\.645±\\pm0\.0830\.625±\\pm0\.089†\\daggerWe are unable to find a Python package that provides an LRP implementation that is compatible with transformer\-based foundation models![Refer to caption](https://arxiv.org/html/2608.13676v1/tuh.jpg)Fig\. 4:EEG\-PRISM results on TUSZ dataset\.Top:Range of spectral attributions aggregated into five bands across all EEG windows in five models\.Bottom:Confusion Matrices \(CM\) comparing region with maximum aggregated spatial attributions of EEG\-PRISM vs ground truth seizure onset zone in clinical notes for five models\.LA: Left Anterior; RA: Right Anterior; LP: Left Posterior; RP: Right Posterior; CS: Central/Subcortical

## IIIResults

We evaluate thepost hocspectral and spatial attributions provided by EEG\-PRISM across several configurations\. Specifically, we train each foundation model on all three datasets for the specific tasks described in Section II\.B\. The models are evaluated on a held\-out test set to verify proper learning\. For completeness, we report the cross\-validated classification performances in Appendix[A](https://arxiv.org/html/2608.13676#A1); however, we note that EEG\-PRISM is agnostic to classification and can be applied toanytrained model\. We use standard explainers to derive the input time\-channel attribution scores on the test data and apply EEG\-PRISM to derive the spectral and spatial attributions\. We use simulated data to quantify the accuracy of the transformed attributions, and subsequently demonstrate the value\-add of EEG\-PRISM for recording\-level signal characterization in the TUSZ dataset and group\-level exploratory biomarkers in the ACE dataset\.

### III\-ASimulated Dataset: Comparison with Ground Truth

The simulated dataset contains known ground\-truth spectral and spatial characteristics associated with each class label, which enables us to quantify the accuracy of the EEG\-PRISM attribution mapping\. We note that the classification task, itself, is trivial for all five foundation models \(Appendix[A](https://arxiv.org/html/2608.13676#A1)\), which provides confidence that the models are learning discriminative representations from the data\.

For the spectral analysis, we compute the total positive attribution within frequency bands corresponding to each predicted class label\. We then compare the band with the maximum attribution to the ground\-truth spectral component used during signal generation\. Since there is a direct association between frequency band and ground\-truth class labels, all four explainability methods achieve near\-perfect accuracy in Table[I](https://arxiv.org/html/2608.13676#S2.T1)\(left\), with minor deviations observed for IG and DeepLift\. DeepSHAP achieves the best performance and consistently recovers the ground\-truth dominant frequency band from attribution scores\. LRP performs comparably on EEGNet and ABNet, but we are unable to find a Python implementation for LRP that is stable for the remaining transformer\-based foundation models\. Overall, EEG\-PRISM can accurately identify the spectral features that are associated with the underlying group differences\. This performance is consistentacross multiple models and explainers\.

For the spatial analysis, we use EEG\-PRISM to map the input attributions onto 5,526 cortical and subcortical sources defined on the MNE fsaverage template\. Although the EEG data was generated using 74,382 sources in the forward model, we reduced the dimensionality for the inverse problem for stability given the ill\-posed nature of this problem\. In contrast to the spectral evaluation, there no association between the source location and class label, as the former is randomly selected for each EEG input\. To quantify the performance of EEG\-PRISM, we first identify the spatial locations with the top 5% of positive attribution scores for a given EEG recording\. We then compute the center of mass for the retained locations and evaluate the the Euclidean distance between this center of mass and the true simulated source location\. These distances are aggregated across the simulated dataset\. Across methods, the average localization error is approximately3±23\\pm 2cm, indicating generally accurate recovery\.

To provide a more nuanced evaluation, we compute quadrant\-level accuracy by first assigning the center of mass to one of five regions: left anterior, right anterior, left posterior, right posterior, and subcortical/central \(chance level = 20%\)\. We then compute the lobe\-level accuracy of this assignment relative to the ground truth\. As shown in Table[I](https://arxiv.org/html/2608.13676#S2.T1), all methods perform above chance in both datasets\. IG achieves the best overall performance, indicating that gradient\-based methods generally outperform others in this setting with approximate inverse solutions\. Among the models, Labram shows better spatial accuracy followed by Cbramod but are not statistically different from each other\. Overall, EEG\-PRISM correctly localizes the attributions to sources that drive EEG, even in a high\-dimensional spatial domain and when those locations are not directly associated with task labels\. This enables robust post\-hoc EEG analysis\.

![Refer to caption](https://arxiv.org/html/2608.13676v1/ace_ext_results.jpg)Fig\. 5:EEG\-PRISM results on ACE dataset grouped into ASD subjects \(top\), TDC subjects \(middle\), and respective differences across two groups \(bottom\)\. Spectral attributions are shown as box plots \(left\), and spatial attributions are mapped onto the 3D surface \(right\)\.
### III\-BTUSZ Dataset: Window\-Level seizure analysis

We investigate whether EEG\-PRISM can accurately reveal the clinically relevant spectral and spatial information associated with an epileptic seizure\. As noted in Section II\.B, the foundation models are trained for window\-level seizure versus non\-seizure classification and do not directly use spectral or spatial information\. We note that the foundation models achieve seizure classification accuracies near88%88\\%\(Appendix[A](https://arxiv.org/html/2608.13676#A1)\), which is on par with the literature\[[54](https://arxiv.org/html/2608.13676#bib.bib54)\]\. After training, we use IG to obtain the time\-channel attribution scores on test sets\. IG was selected due to its computational efficiency, applicability across all models, and consistent performance in the simulation study\.

In the spectral domain, we selected the correctly classified seizure windows in each EEG recording and averaged the positive EEG\-PRISM spectral attributions across the five standard frequency bands\. We do not include negative attributions, as they are inversely related to the predicted class\. As shown in Fig\.[4](https://arxiv.org/html/2608.13676#S2.F4)\(top\), delta and theta bands consistently exhibited higher importance, with theta being dominant in three of the five models\. These findings align with clinical annotations: 43% of subject reports abnormalities in the theta band and 35% also mention the delta band\. These bands are commonly associated with epileptiform activity such as spike\-sharp waves following by slow waves\[[55](https://arxiv.org/html/2608.13676#bib.bib55)\]\. In contrast, the alpha and beta bands were referenced in 12% and 11% of clinical reports, respectively, and received lower attribution scores in that order\. The gamma band showed minimal importance, likely due to the preprocessing steps that filtered out frequencies above 30 Hz\. In summary, the spectral patterns uncovered by EEG\-PRISM are consistent with clinical reports across foundation models and can be leveraged for early prediction, seizure tracking, and clinical stratification\.

In the spatial domain, we analyzed the first 15 windows \(30 seconds\) of seizure activity for each recording to determine whether EEG\-PRISM localize seizure onset zone \(SOZ\)\. Here, we used EEG\-PRISM to obtain attribution scores over 5,524 cortical and subcortical sources\. Similar to the simulated experiment, we aggregated the positive attributions into five regions: left/right anterior, left/right posterior, and central\. The region with the highest attribution was taken as the predicted SOZ and compared against clinical ground truth obtained from the report\. As shown in Fig\.[4](https://arxiv.org/html/2608.13676#S2.F4)\(bottom\), the resulting confusion matrix exhibits a diagonal\-dominant structure, which indicates meaningful localization performance\. We observe that central regions were the greatest source of confusion, likely due to the difficulty in estimating subcortical sources from EEG\. Another source of error was confusion between the right and left hemispheres, which may be attributed to strong inter\-hemispheric connections in the brain\. Overall, EEG\-PRISM achieved an average localization accuracy of roughly 50% across all models, which is comparable tosupervised methodsfor SOZ localization\[[56](https://arxiv.org/html/2608.13676#bib.bib56)\]\. Notably, this performance is achieved without any prior knowledge of the SOZ when training the foundation models\. Furthermore, EEG\-PRISM enables SOZ localization in 3\-D source space, which provides a finer spatial resolution for presurgical planning in epilepsy\.

### III\-CACE Dataset: Group\-Level Autism Analysis

Our final experiment demonstrates how EEG\-PRISM can be paired with foundation models for exploratory group\-level biomarker analyses\. We consider the problem of ASD versus TDC classification\. Here, the foundation models perform a binary classification for each 2\-second EEG window and then aggregate the predictions to the subject level via majority voting across all windows\. The LaBraM foundation model achieves the highest classification accuracy \(Appendix[A](https://arxiv.org/html/2608.13676#A1), Table[III](https://arxiv.org/html/2608.13676#A1.T3)\) and provides the backbone for EEG\-PRISM\. We choose best\-performing models to ensure reliable interpretation of attribution scores; however, model performance does not dictate whether EEG\-PRISM can be applied, as it operates post\-hoc\. Once again, we rely on IG to obtain the original time\-channel attributions\.

For the spectral analysis, we aggregate the EEG\-PRISM attributions into the five canonical frequency bands and average the positive attribution scores across EEG windows for a given subject\. For the spatial analysis, we visualize 5,526 source\-level attribution scores, averaged across subjects within each group\. For visualization, we interpolate the scores using the NiLearn and PyVista packages\[[57](https://arxiv.org/html/2608.13676#bib.bib57),[58](https://arxiv.org/html/2608.13676#bib.bib58)\]\. Similar to TUSZ, we restrict the analysis to correctly classified subjects and time windows for reliable interpretation\.

As shown in Fig\.[5](https://arxiv.org/html/2608.13676#S3.F5)\(left\), correct ASD classification is primarily associated with increased attributions in the delta frequency band, whereas correct TDC classification is driven by alpha band activity\. In the context of ASD, which is characterized by differences in social communication, increased delta activity is thought to reflect altered cortical processing, while disrupted alpha rhythms indicate differences in attention and cognitive engagement\. Prior studies report changes in both delta power and alpha connectivity in ASD consistent with these patterns\[[59](https://arxiv.org/html/2608.13676#bib.bib59),[60](https://arxiv.org/html/2608.13676#bib.bib60)\]\. At the spatial level, both the ASD and TDC groups exhibit similar attribution patterns, with frontal lobe driving the classification\. The group\-level differences in spatial attributions tend to localize in the temporal lobes and extend into right central regions\. Interestingly, this interpretation is consistent with established literature: prefrontal cortex supports working memory and higher cognitive processing\[[61](https://arxiv.org/html/2608.13676#bib.bib61)\], while the temporal lobes contribute to auditory processing, language, and social communication\[[62](https://arxiv.org/html/2608.13676#bib.bib62)\]\. Thus, EEG\-PRISM enables post\-hoc biomarker discovery with foundation models consistent with prior work\.

## IVDiscussion

In this work, we addressed a key limitation of emerging foundation models for EEG analysis\. While these models achieve strong performance using high\-dimensional time–channel inputs, they are opaque with limited interpretability\. Post\-hoc “explainers" partially mitigate this issue by identifying the input features associated with a given prediction\. However, the input EEG \(i\.e\., time\-channel space\) does not align with clinically meaningful spectral and spatial representations\. Therefore, we have introduced EEG\-PRISM as a principled framework to project the input\-level attribution scores derived frommulitple post\-hoc explainersinto other domains\. EEG\-PRISM uses a combination of signal processing theory with chain rule\-based backpropagation to enable the mapping to be donewithout modifying or retraining the foundation model\. We have systematically evaluated the accuracy and fidelity of EEG\-PRISM across five foundation models and four attribution methods\. We have further validated its clinical utility in epilepsy by demonstrating its alignment with spectral and spatial patterns in clinical reports\. Finally, we have demonstrated the potential of EEG\-PRISM for group\-level biomarker analysis using autism as a case study\. Taken together, EEG\-PRISM is a lightweight and ubiquitous tool that seamlessly works across foundation models, explainability methods, and clinical/research domains\.

Intuitively, mapping the time\-channel attributions to the spectral domain is straightforward due to the invertible DFT incorporated within EEG\-PRISM\. In simulated experiments, this mapping capitalizes on the association between frequency band and class label to achieve near\-perfect recovery of the ground truth phenomenon\. In epilepsy, EEG\-PRISM consistently identified delta–theta contributions to seizure detection, which aligns with the established clinical knowledge of epileptiform activity\[[55](https://arxiv.org/html/2608.13676#bib.bib55),[63](https://arxiv.org/html/2608.13676#bib.bib63)\]\. In our exploratory autism analysis, EEG\-PRISM identified interactions between delta and alpha bands as influencing model predictions\. Delta slowing is commonly associated with atypical cortical processing, while alpha rhythms are linked to wakeful rest, attention, and cognitive engagement\. Prior studies in autism have reported alterations in both delta power and alpha connectivity in autism\[[59](https://arxiv.org/html/2608.13676#bib.bib59),[60](https://arxiv.org/html/2608.13676#bib.bib60)\]\. More broadly, spectral EEG features are closely tied to brain function: delta activity relates to sleep and pathological slowing, theta to memory and cognitive control, alpha to attention and inhibitory processing, beta to motor activity, and gamma to higher\-order cognition and binding\[[64](https://arxiv.org/html/2608.13676#bib.bib64)\]\. EEG\-PRISM enables any foundation model to be interpreted in this space, ensuring that model explanations remain aligned with established neuroscientific principles and clinical patterns\.

The mapping from input to spatial source domain using EEG\-PRISM relies on user\-defined forward and inverse models\. Despite this challenge, it achieved high accuracy in simulated experiments with known ground truth phenomena, thus demonstrating robustpost\-hoc recoveryof spatial patterns\. EEG\-PRISM also achieved competitive SOZ localization performance in a fully post hoc setting, despite the foundation models not having access to spatial supervision during training\. In fact, it performed comparably to prior supervised approaches for SOZ localization\[[56](https://arxiv.org/html/2608.13676#bib.bib56)\]\. These results highlight the potential of EEG\-PRISM for clinical translation, including applications in treatment planning for epilepsy\. Extending this framework to autism, we identified neurologically plausible associations in the frontal and temporal areas\. The prefrontal cortex is widely implicated in cognitive control and working memory, while temporal and central regions are associated with memory, auditory processing, language, and emotional regulation which likely differ between individuals with ASD and TDCs\[[65](https://arxiv.org/html/2608.13676#bib.bib65),[66](https://arxiv.org/html/2608.13676#bib.bib66)\]\. Thus, EEG\-PRISM can identify spatially meaningful regions that drive model predictions, thus providing an important foundation for noninvasive biomarker discovery and hypothesis generation in neuropsychiatric disorders\.

Another way to evaluate the “information gain" of EEG\-PRISM is via the mean\-normalized range of the attribution scores in each domain\. Specifically, in cases where feature strongly influences the output, the attribution scores will be uniformly distributed, leading to low range\. Conversely, if there are a few dominant features, then the attribution scores will be more concentrated and increase the range\. Mean normalization ensures that the values are comparable across vectors of different sizes\. Table[II](https://arxiv.org/html/2608.13676#S4.T2)reports the mean\-normalized ranges for the Labram model and IG explainer in the original time\-channel domain, the spectral domain, and the spatial source domain\. The default time\-channel domain exhibits the lowest mean\-normalized range, indicating that it isnotthe natural space to learn predictive representations\. Across datasets, the spectral domain shows substantially higher mean\-normalized range\. The spatial domain also shows higher values than input space, particularly in ACE dataset where we saw that the information was localized to frontal regions\. Overall, our analysis suggests that there is more concentrated information in the spectral and spatial domains, as often used in clinical practice\.

TABLE II:mean\-normalized range of EEG\-PRISM attributionsIn this work, we applied EEG\-PRISM to a variety of foundation models and explainers to demonstrate its universal applicability\. Currently, larger pretrained models, such as Labram, Cbramod, and LUNA, provide a strong starting point and adapt quickly with small datasets like ACE\. For real\-world datasets with well\-understood tasks \(e\.g\., seizure detection\), model choice matters less, i\.e\., most models perform similarly\. Across models, attribution scores are broadly consistent, though we expect that the best model/explainer configuration will differ across real\-world datasets\. Our work demonstrates that regardless of the foundation model and downstream task, we can use EEG\-PRISM to obtain more clinically relevant interpretations\.

Likewise, we have shown that EEG\-PRISM pairs well with a range of backpropagation\-based attribution methods\. Notably, IG performs well on models with smooth gradients, has strong axiomatic properties, and produces consistent results\. DeepLIFT addresses gradient saturation using discrete differences and model\-specific rules\. However, it can show instability with a single baseline\. DeepSHAP approximates Shapley values by combining DeepLIFT with multiple baselines\. It performed best for deterministic spectral attributions\. However, it showed slight degradation in the spatial domain, likely due to error magnification across multiple baselines\. As foundation models increasingly adopt transformer architectures, EEG\-PRISM with IG presents as a stable, model\-agnostic attribution approach across datasets\. LRP was not applied to transformers in this study, as LRP rules for self\-attention remain an active area of research\.

One limitation of this work is that we analyzed spectral and source domains separately\. Future work should explore joint spectral\-source representations as a methodological extension while ensuring that approximate inversion does not compromise spectral interpretation\. Another limitation is that our theoretical guarantees were derived for linear mappings only\. Future work will explore mappings for complex nonlinear transformations, thus allowing researchers to investigate EEG connectivity and their relationship with model outputs\. Approaches such as linear surrogate models and autoencoder\-based inversion may provide practical strategies for approximating these transformations\. Finally, using a simple average to summarize attribution patterns across EEG windows can obscure clinically\-relevant signal variability\. Thus, we will explore algorithmic improvements to identify consistent attributions across windows and across subjects for group\-level analyses\. Future work will apply EEG\-PRISM to other neurological conditions, including disease subtyping and treatment response monitoring by tracking identified multi\-domain biomarkers\.

## VConclusion

In summary, we presented EEG\-PRISM as a principled and flexible framework for interpreting EEG foundation models by projecting attribution scores into clinically meaningful spectral and spatial domains\. Across simulated and real\-world datasets, EEG\-PRISM demonstrated strong fidelity in recovering relevant frequency patterns and spatial regions, aligning with established neuroscientific and clinical knowledge\. Despite operating in a fully post hoc and model\-agnostic setting, it achieved competitive performance in tasks such as seizure characterization and localization, while also showing promise in exploratory analyses such as autism\. By bridging the gap between high\-performing black\-box models and interpretable neurological insights, EEG\-PRISM provides a practical pathway toward more transparent, reliable, and clinically grounded EEG analysis, with broad potential for future research and translational applications\.

## Acknowledgment

AI tools assisted with general code completion and manuscript formatting\. All content were thoroughly reviewed by the authors\. We thank the ACE Network and GENDAAR for collecting EEG data\.

## Appendix AFoundation Model Performance on the Primary Prediction Task

Table[III](https://arxiv.org/html/2608.13676#A1.T3)reports the average classification accuracy of the primary task of five foundation models in simulated, TUSZ, and ACE datasets\. In simulated and TUSZ datasets, window\-level accuracy of abnormality \(4\-way\) and seizure classification \(binary\) are reported, respectively, where all models achieved similar performance\. In ACE dataset, subject level accuracy in ASD vs TDC binary classification is reported after majority voting of window\-level model output across subject’s EEG recording\. Labram achieved the the highest accuracy \(66%\), followed by LUNA \(62%\)\.

TABLE III:model accuracy in classification across datasets
## References

- \[1\]B\. Tharp and W\. Gersch, “Spectral analysis of seizures in humans,”*Computers and Biomedical Research*, vol\. 8, no\. 6, pp\. 503–521, 1975\.
- \[2\]A\. Goenka, A\. Boro, and E\. Yozawitz, “Comparative sensitivity of quantitative eeg \(qeeg\) spectrograms for detecting seizure subtypes,”*Seizure*, vol\. 55, pp\. 70–75, 2018\.
- \[3\]H\. P\. Zaveri, R\. B\. Duckrow, N\. C\. De Lanerolle, and S\. S\. Spencer, “Distinguishing subtypes of temporal lobe epilepsy with background hippocampal activity,”*Epilepsia*, vol\. 42, no\. 6, pp\. 725–730, 2001\.
- \[4\]C\.\-S\. Ouyang, C\.\-T\. Chiang, R\.\-C\. Yang, R\.\-C\. Wu, H\.\-C\. Wu, and L\.\-C\. Lin, “Quantitative eeg findings and response to treatment with antiepileptic medications in children with epilepsy,”*Brain and Development*, vol\. 40, no\. 1, pp\. 26–35, 2018\.
- \[5\]J\. R\. Hughes and E\. R\. John, “Conventional and quantitative electroencephalography in psychiatry,”*The Journal of neuropsychiatry and clinical neurosciences*, vol\. 11, no\. 2, pp\. 190–208, 1999\.
- \[6\]J\. J\. Newson and T\. C\. Thiagarajan, “Eeg frequency bands in psychiatric disorders: a review of resting state studies,”*Frontiers in human neuroscience*, vol\. 12, p\. 521, 2019\.
- \[7\]W\. S\. Neo, D\. Foti, B\. Keehn, and B\. Kelleher, “Resting\-state eeg power differences in autism spectrum disorder: a systematic review and meta\-analysis,”*Translational Psychiatry*, vol\. 13, no\. 1, p\. 389, 2023\.
- \[8\]H\. Hallez, B\. Vanrumste, R\. Grech, J\. Muscat, W\. De Clercq, A\. Vergult, Y\. D’Asseler, K\. P\. Camilleri, S\. G\. Fabri, S\. Van Huffel*et al\.*, “Review on solving the forward problem in eeg source analysis,”*Journal of neuroengineering and rehabilitation*, vol\. 4, no\. 1, p\. 46, 2007\.
- \[9\]A\. Ray, J\. X\. Tao, S\. M\. Hawes\-Ebersole, and J\. S\. Ebersole, “Localizing value of scalp eeg spikes: a simultaneous scalp and intracranial study,”*Clinical neurophysiology*, vol\. 118, no\. 1, pp\. 69–79, 2007\.
- \[10\]K\. L\. Coburn, E\. C\. Lauterbach, N\. N\. Boutros, K\. J\. Black, D\. B\. Arciniegas, and C\. E\. Coffey, “The value of quantitative electroencephalography in clinical psychiatry: a report by the committee on research of the american neuropsychiatric association,”*The Journal of neuropsychiatry and clinical neurosciences*, vol\. 18, no\. 4, pp\. 460–500, 2006\.
- \[11\]R\. Sharma and H\. K\. Meena, “Emerging trends in eeg signal processing: A systematic review,”*SN Computer Science*, vol\. 5, no\. 4, p\. 415, 2024\.
- \[12\]Y\. Zhao, L\. Cao, Y\. Ji, B\. Wang, and W\. Wu, “Interpretable eeg emotion classification via cnn model and gradient\-weighted class activation mapping,”*Brain Sciences*, vol\. 15, no\. 8, p\. 886, 2025\.
- \[13\]S\. Sylvester, M\. Sagehorn, T\. Gruber, M\. Atzmueller, and B\. Schöne, “Shap value\-based erp analysis \(sherpa\): Increasing the sensitivity of eeg signals with explainable ai methods,”*Behavior Research Methods*, vol\. 56, no\. 6, pp\. 6067–6081, 2024\.
- \[14\]A\. K\. Singh and S\. Krishnan, “Trends in eeg signal feature extraction applications,”*Frontiers in artificial intelligence*, vol\. 5, p\. 1072801, 2023\.
- \[15\]K\. Hornik, “Approximation capabilities of multilayer feedforward networks,”*Neural networks*, vol\. 4, no\. 2, pp\. 251–257, 1991\.
- \[16\]G\. Kuruppu, N\. Wagh, V\. Kremen, and Y\. Varatharajah, “Eeg foundation models: a critical review of current progress and future directions,”*Journal of neural engineering*, vol\. 23, no\. 2, p\. 021001, 2026\.
- \[17\]V\. J\. Lawhern, A\. J\. Solon, N\. R\. Waytowich, S\. M\. Gordon, C\. P\. Hung, and B\. J\. Lance, “Eegnet: a compact convolutional neural network for eeg\-based brain–computer interfaces,”*Journal of neural engineering*, vol\. 15, no\. 5, p\. 056013, 2018\.
- \[18\]Y\. Song, Q\. Zheng, B\. Liu, and X\. Gao, “Eeg conformer: Convolutional transformer for eeg decoding and visualization,”*IEEE Transactions on Neural Systems and Rehabilitation Engineering*, vol\. 31, pp\. 710–719, 2022\.
- \[19\]D\. Kostas, S\. Aroca\-Ouellette, and F\. Rudzicz, “Bendr: Using transformers and a contrastive self\-supervised learning task to learn from massive amounts of eeg data,”*Frontiers in Human Neuroscience*, vol\. 15, p\. 653659, 2021\.
- \[20\]C\. Yang, M\. Westover, and J\. Sun, “Biot: Biosignal transformer for cross\-data learning in the wild,”*Advances in Neural Information Processing Systems*, vol\. 36, pp\. 78 240–78 260, 2023\.
- \[21\]W\.\-B\. Jiang, L\. Zhao, and B\.\-L\. Lu, “Large brain model for learning generic representations with tremendous eeg data in bci,” in*International Conference on Learning Representations*, vol\. 2024, 2024, pp\. 16 405–16 426\.
- \[22\]E\. Shi, K\. Zhao, Q\. Yuan, J\. Wang, H\. Hu, S\. Yu, and S\. Zhang, “Fome: A foundation model for eeg using adaptive temporal\-lateral attention scaling,”*arXiv preprint arXiv:2409\.12454*, 2024\.
- \[23\]J\. Wang, S\. Zhao, Z\. Luo, Y\. Zhou, H\. Jiang, S\. Li, T\. Li, and G\. Pan, “Cbramod: A criss\-cross brain foundation model for eeg decoding,” in*International conference on learning representations*, vol\. 2025, 2025, pp\. 75 310–75 346\.
- \[24\]B\. Döner, T\. M\. Ingolfsson, L\. Benini, and Y\. Li, “Luna: Efficient and topology\-agnostic foundation model for eeg signal analysis,”*Advances in Neural Information Processing Systems*, vol\. 38, pp\. 70 682–70 708, 2026\.
- \[25\]O\. Presacan, J\. Ojha, A\. Yazidi, E\. Monteiro, and P\. G\. Lind, “A comprehensive review of explainable ai in deep learning algorithms for eeg analysis,”*ACM Transactions on Computing for Healthcare*, vol\. 7, no\. 2, pp\. 1–28, 2026\.
- \[26\]M\. C\. Maurer, J\. M\. Metsch, P\. Hempel, T\. Bender, N\. Spicher, and A\.\-C\. Hauschild, “Explainable artificial intelligence on biosignals for clinical decision support,” in*Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining*, 2024, pp\. 6597–6604\.
- \[27\]G\. Montavon, A\. Binder, S\. Lapuschkin, W\. Samek, and K\.\-R\. Müller, “Layer\-wise relevance propagation: an overview,”*Explainable AI: interpreting, explaining and visualizing deep learning*, pp\. 193–209, 2019\.
- \[28\]R\. Achtibat, S\. M\. V\. Hatefi, M\. Dreyer, A\. Jain, T\. Wiegand, S\. Lapuschkin, and W\. Samek, “Attnlrp: attention\-aware layer\-wise relevance propagation for transformers,”*arXiv preprint arXiv:2402\.05602*, 2024\.
- \[29\]J\. M\. z\. Bexten, N\. Scherf, B\. Franczyk, and S\. M\. Hofmann, “From clever hans to scientific discovery: Interpreting eeg foundational transformers with lrp,”*arXiv preprint arXiv:2605\.11885*, 2026\.
- \[30\]H\. Nam, J\.\-M\. Kim, W\. Choi, S\. Bak, and T\.\-E\. Kam, “The effects of layer\-wise relevance propagation\-based feature selection for eeg classification: a comparative study on multiple datasets,”*Frontiers in Human Neuroscience*, vol\. 17, p\. 1205881, 2023\.
- \[31\]A\. Nouri and Z\. Tabanfar, “Detection of adhd disorder in children using layer\-wise relevance propagation and convolutional neural network: An eeg analysis,”*Frontiers in Biomedical Technologies*, vol\. 11, no\. 1, pp\. 14–21, 2024\.
- \[32\]M\. Sundararajan, A\. Taly, and Q\. Yan, “Axiomatic attribution for deep networks,” in*International conference on machine learning*\. PMLR, 2017, pp\. 3319–3328\.
- \[33\]D\. Maiti, M\. Basak, and S\. Hazra, “Neuroexplain: An explainable deep learning framework for multiclass emotion recognition from gameplay eeg,” in*2025 5th International Conference on Emerging Research in Electronics, Computer Science and Technology \(ICERECT\)*\. IEEE, 2025, pp\. 1–6\.
- \[34\]K\. Tachikawa, Y\. Kawai, J\. Park, and M\. Asada, “Compensated integrated gradients to reliably interpret eeg classification,”*arXiv preprint arXiv:1811\.08633*, 2018\.
- \[35\]A\. Shrikumar, P\. Greenside, and A\. Kundaje, “Learning important features through propagating activation differences,” in*International conference on machine learning*\. PMlR, 2017, pp\. 3145–3153\.
- \[36\]D\. Park, H\. Park, S\. Kim, S\. Choo, S\. Lee, C\. S\. Nam, and J\.\-Y\. Jung, “Spatio\-temporal explanation of 3d\-eegnet for motor imagery eeg classification using permutation and saliency,”*IEEE Transactions on Neural Systems and Rehabilitation Engineering*, vol\. 31, pp\. 4504–4513, 2023\.
- \[37\]A\. Sujatha Ravindran and J\. Contreras\-Vidal, “An empirical comparison of deep learning explainability approaches for eeg using simulated ground truth,”*Scientific Reports*, vol\. 13, no\. 1, p\. 17709, 2023\.
- \[38\]S\. M\. Lundberg and S\.\-I\. Lee, “A unified approach to interpreting model predictions,”*Advances in neural information processing systems*, vol\. 30, 2017\.
- \[39\]T\. Shawly and A\. A\. Alsheikhy, “Eeg\-based detection of epileptic seizures in patients with disabilities using a novel attention\-driven deep learning framework with shap interpretability,”*Egyptian Informatics Journal*, vol\. 31, p\. 100734, 2025\.
- \[40\]B\. Mouazen, A\. Bendaouia, O\. Bellakhdar, K\. Laghdaf, A\. Ennair, E\. H\. Abdelwahed, and G\. De Marco, “Transparent eeg analysis: leveraging autoencoders, bi\-lstms, and shap for improved neurodegenerative diseases detection,”*Sensors*, vol\. 25, no\. 18, p\. 5690, 2025\.
- \[41\]A\. Almadhor, S\. Ojo, T\. I\. Nathaniel, S\. Alsubai, A\. Alharthi, A\. A\. Hejaili, and G\. A\. Sampedro, “An interpretable xai deep eeg model for schizophrenia diagnosis using feature selection and attention mechanisms,”*Frontiers in Oncology*, vol\. 15, p\. 1630291, 2025\.
- \[42\]J\. Vielhaben, S\. Lapuschkin, G\. Montavon, and W\. Samek, “Explainable ai for time series via virtual inspection layers,”*Pattern Recognition*, vol\. 150, p\. 110309, 2024\.
- \[43\]L\. R\. Krol, J\. Pawlitzki, F\. Lotte, K\. Gramann, and T\. O\. Zander, “Sereega: Simulating event\-related eeg activity,”*Journal of neuroscience methods*, vol\. 309, pp\. 13–24, 2018\.
- \[44\]Y\. Huang, L\. C\. Parra, and S\. Haufe, “The new york head‚Äîa precise standardized volume conductor model for eeg source localization and tes targeting,”*NeuroImage*, vol\. 140, pp\. 150–162, 2016\.
- \[45\]V\. Shah, E\. Von Weltin, S\. Lopez, J\. R\. McHugh, L\. Veloso, M\. Golmohammadi, I\. Obeid, and J\. Picone, “The temple university hospital seizure detection corpus,”*Frontiers in neuroinformatics*, vol\. 12, p\. 83, 2018\.
- \[46\]E\. Neuhaus, S\. J\. Lowry, M\. Santhosh, A\. Kresse, L\. A\. Edwards, J\. Keller, E\. J\. Libsack, V\. Y\. Kang, A\. Naples, A\. Jack*et al\.*, “Resting state eeg in youth with asd: age, sex, and relation to phenotype,”*Journal of neurodevelopmental disorders*, vol\. 13, pp\. 1–15, 2021\.
- \[47\]M\. Wimpff, L\. Gizzi, J\. Zerfowski, and B\. Yang, “Eeg motor imagery decoding: A framework for comparative analysis with channel attention mechanisms,”*Journal of Neural Engineering*, 2024\.
- \[48\]B\. Aristimunha, P\. Guetschel, M\. Wimpff, L\. Gemein, C\. Rommel, H\. Banville, M\. Sliwowski, D\. Wilson, S\. Brandt, T\. Gnassounou, J\. Paillard, B\. Junqueira Lopes, S\. Sedlar, T\. Moreau, S\. Chevallier, A\. Gramfort, and R\. T\. Schirrmeister, “Braindecode: Toolbox for decoding raw electrophysiological brain data with deep learning models,” 2025,[https://github\.com/braindecode/braindecode](https://github.com/braindecode/braindecode)\.
- \[49\]D\. P\. Kingma and J\. Ba, “Adam: A method for stochastic optimization,”*arXiv preprint arXiv:1412\.6980*, 2014\.
- \[50\]N\. Kokhlikyan, V\. Miglani, M\. Martin, E\. Wang, B\. Alsallakh, J\. Reynolds, A\. Melnikov, N\. Kliushkina, C\. Araya, S\. Yan*et al\.*, “Captum: A unified and generic model interpretability library for pytorch,”*arXiv preprint arXiv:2009\.07896*, 2020\.
- \[51\]A\. Gramfort, M\. Luessi, E\. Larson, D\. A\. Engemann, D\. Strohmeier, C\. Brodbeck, R\. Goj, M\. Jas, T\. Brooks, L\. Parkkonen*et al\.*, “Meg and eeg data analysis with mne\-python,”*Frontiers in Neuroinformatics*, vol\. 7, p\. 267, 2013\.
- \[52\]R\. D\. Pascual\-Marqui*et al\.*, “Standardized low\-resolution brain electromagnetic tomography \(sloreta\): technical details,”*Methods find exp clin pharmacol*, vol\. 24, no\. Suppl D, pp\. 5–12, 2002\.
- \[53\]R\. Grech, T\. Cassar, J\. Muscat, K\. P\. Camilleri, S\. G\. Fabri, M\. Zervakis, P\. Xanthopoulos, V\. Sakkalis, and B\. Vanrumste, “Review on solving the inverse problem in eeg source analysis,”*Journal of neuroengineering and rehabilitation*, vol\. 5, no\. 1, p\. 25, 2008\.
- \[54\]M\. Zabihi, E\. J\. Gilmore, K\. Ding, and E\. S\. Rosenthal, “A transparent ai assurance and benchmarking framework for eeg seizure detection on tusz seeded with a reproducible gradient\-boosting ensemble,”*Scientific Reports*, vol\. 16, no\. 1, p\. 11283, 2026\.
- \[55\]P\. D\. Emmady, R\. M\. D\. Asuncion, and A\. C\. Anilkumar, “Eeg abnormal waveforms,”*StatPearls \[internet\]*, 2025\.
- \[56\]D\. M\. Shama, J\. Jing, and A\. Venkataraman, “Deepsoz: A robust deep model for joint temporal and spatial seizure onset localization from multichannel eeg data,” in*International Conference on Medical Image Computing and Computer\-Assisted Intervention*\. Springer, 2023, pp\. 184–194\.
- \[57\]A\. Abraham, F\. Pedregosa, M\. Eickenberg, P\. Gervais, A\. Mueller, J\. Kossaifi, A\. Gramfort, B\. Thirion, and G\. Varoquaux, “Machine learning for neuroimaging with scikit\-learn,”*Frontiers in neuroinformatics*, vol\. 8, p\. 14, 2014\.
- \[58\]C\. Sullivan and A\. Kaszynski, “Pyvista: 3d plotting and mesh analysis through a streamlined interface for the visualization toolkit \(vtk\),”*Journal of Open Source Software*, vol\. 4, no\. 37, p\. 1450, 2019\.
- \[59\]J\. Wang, J\. Barstein, L\. E\. Ethridge, M\. W\. Mosconi, Y\. Takarae, and J\. A\. Sweeney, “Resting state eeg abnormalities in autism spectrum disorders,”*Journal of neurodevelopmental disorders*, vol\. 5, no\. 1, p\. 24, 2013\.
- \[60\]M\. Kopańska, D\. Ochojska, I\. Sarzyńska, J\. Trojniak, and J\. Szczygielski, “Exploratory quantitative eeg characteristics in children with autism spectrum disorder,”*Frontiers in Psychiatry*, vol\. 16, p\. 1689000, 2025\.
- \[61\]R\. M\. El\-Baba and M\. P\. Schury, “Neuroanatomy, frontal cortex,” 2023, statPearls \[Internet\]\. Treasure Island \(FL\): StatPearls Publishing\. Updated May 29, 2023\.[https://www\.ncbi\.nlm\.nih\.gov/books/NBK554483/](https://www.ncbi.nlm.nih.gov/books/NBK554483/)\.
- \[62\]M\. A\. Patriquin, T\. DeRamus, L\. E\. Libero, A\. Laird, and R\. K\. Kana, “Neuroanatomical and neurofunctional markers of social cognition in autism spectrum disorder,”*Human brain mapping*, vol\. 37, no\. 11, pp\. 3957–3978, 2016\.
- \[63\]B\. Clemens, G\. Szigeti, and Z\. Barta, “Eeg frequency profiles of idiopathic generalised epilepsy syndromes,”*Epilepsy research*, vol\. 42, no\. 2\-3, pp\. 105–115, 2000\.
- \[64\]G\. Buzsáki,*Rhythms of the Brain*\. Oxford university press, 2006\.
- \[65\]K\. A\. Pelphrey and E\. J\. Carter, “Brain mechanisms for social perception: lessons from autism and typical development,”*Annals of the New York Academy of Sciences*, vol\. 1145, no\. 1, pp\. 283–299, 2008\.
- \[66\]E\. Courchesne, K\. Pierce, C\. M\. Schumann, E\. Redcay, J\. A\. Buckwalter, D\. P\. Kennedy, and J\. Morgan, “Mapping early brain development in autism,”*Neuron*, vol\. 56, no\. 2, pp\. 399–413, 2007\.