MEL: Coordinate-Preserving EEG Tokenization for fMRI Translation

# MEL: Coordinate-Preserving EEG Tokenization for fMRI Translation
Source: [https://arxiv.org/html/2608.29304](https://arxiv.org/html/2608.29304)
###### Abstract

Translating electroencephalography \(EEG\) into functional magnetic resonance imaging \(fMRI\) is important for medical neuroimaging, clinical brain\-state monitoring, and multimodal neural decoding, because it aims to infer spatially organized hemodynamic activity from fast and accessible electrophysiological recordings\. Existing EEG\-to\-fMRI studies mainly pursue stronger decoders, but the problem is also constrained by a representation\-interface mismatch: fMRI responses are delayed, temporally integrated, and spatially distributed, whereas generic EEG encodings often entangle temporal lag, channel identity, and frequency\-band structure\. We proposeMulti\-band EEG Latent\-state Tokenization \(MEL\), a coordinate\-preserving EEG representation framework that anchors each target fMRI response to its preceding EEG history and organizes it into lag\-channel\-frequency neural\-state tokens\. By explicitly capturing hemodynamic latency and spectral\-spatial dynamics, MEL aligns fMRI\-pertinent EEG representations with capacity\-controlled readouts without depending entirely on model scaling\. Experiments on VU EEG\-fMRI benchmarks and external Oddball data show that MEL improves prediction over strong NeuroBOLT baselines\. Ablations and controls further indicate that the gains come from structured EEG representation rather than leakage, shortcut statistics, or decoder capacity\.

1Beijing Normal\-Hong Kong Baptist University

2The University of Hong Kong

3Peking University

t330034034@mail\.bnbu\.edu\.cn, zetingyan@connect\.hku\.hk, 2300012924@stu\.pku\.edu\.cn,

2101112018@stu\.pku\.edu\.cn, xi\.zhang@pku\.edu\.cn

## Introduction

EEG\-to\-fMRI translation aims to predict spatially distributed blood\-oxygen\-level\-dependent \(BOLD\) responses from electroencephalography \(EEG\)\. EEG provides millisecond\-level temporal resolution of electrophysiological activity, whereas functional magnetic resonance imaging \(fMRI\) offers substantially finer spatial localization across cortical and subcortical regions\. A reliable mapping between these modalities could support multimodal neural decoding, brain\-state monitoring, and the study of large\-scale neural dynamics\. However, EEG\-to\-fMRI prediction is not ordinary cross\-modal regression\. First, BOLD activity at timettreflects neural activity accumulated over a preceding temporal window rather than an instantaneous EEG sample\([Logothetis et al\. 2001](https://arxiv.org/html/2608.29304#bib.bib33);[Buxton et al\. 1998](https://arxiv.org/html/2608.29304#bib.bib4);[Glover 1999](https://arxiv.org/html/2608.29304#bib.bib18)\)\. Second, relevant electrophysiological evidence is distributed across electrodes\. Third, neural activity is expressed through frequency\-specific rhythms whose relationship with BOLD varies across states and anatomical regions\([Buzsáki and Draguhn 2004](https://arxiv.org/html/2608.29304#bib.bib5);[Cohen 2014](https://arxiv.org/html/2608.29304#bib.bib8);[Scheeringa et al\. 2011](https://arxiv.org/html/2608.29304#bib.bib41)\)\. An fMRI target is therefore better viewed as the outcome of a structured neural history over temporal lag, channel, and frequency coordinates\. This leads to the central question:*in what form should electrophysiological histories be represented to preserve the information needed for decoding delayed hemodynamic responses?*When these factors are presented without explicit target\-relative semantics, the decoder must infer their relationships from limited paired EEG\-fMRI data\. Prediction error may therefore arise from a representation\-interface limitation, rather than insufficient decoder capacity alone\.

![Refer to caption](https://arxiv.org/html/2608.29304v1/figure1_mel_motivation.png)Figure 1:Representation\-centric view of ROI\-level EEG\-to\-fMRI prediction\. Many EEG encodings model lag, channel, and frequency information implicitly, whereas MEL exposes these axes as explicit target\-relative coordinates before decoding\.Prior EEG\-to\-fMRI studies approach this mapping through cross\-modal regression, electrode\-graph modeling, transformer\-based transcoding, pretrained EEG encoders, and volumetric generation\([Calhas and Henriques 2020](https://arxiv.org/html/2608.29304#bib.bib6);[Liu and Sajda 2023](https://arxiv.org/html/2608.29304#bib.bib32);[Afrasiyabi et al\. 2025](https://arxiv.org/html/2608.29304#bib.bib2);[Calhas and Henriques 2022](https://arxiv.org/html/2608.29304#bib.bib7);[Lanzino et al\. 2024](https://arxiv.org/html/2608.29304#bib.bib27);[Li et al\. 2024b](https://arxiv.org/html/2608.29304#bib.bib31);[Roos et al\. 2025](https://arxiv.org/html/2608.29304#bib.bib39);[He et al\. 2025](https://arxiv.org/html/2608.29304#bib.bib20);[Yao et al\. 2025](https://arxiv.org/html/2608.29304#bib.bib48)\)\. These methods substantially increase the expressive power of the predictor and can model temporal, spatial, and spectral dependencies\. However, the target\-relative meanings of lag, channel, and frequency are typically learned implicitly within the network\. Meanwhile, simultaneous EEG\-fMRI studies have established delayed neurovascular coupling and systematic associations between band\-limited EEG activity and regional BOLD fluctuations\([Goldman et al\. 2002](https://arxiv.org/html/2608.29304#bib.bib19);[Huster et al\. 2012](https://arxiv.org/html/2608.29304#bib.bib22);[Jorge et al\. 2014](https://arxiv.org/html/2608.29304#bib.bib24);[Philiastides et al\. 2021](https://arxiv.org/html/2608.29304#bib.bib38)\)\.

This distinction is important because preserving information is not equivalent to exposing it in a form that can be reliably learned from limited paired observations\. A flattened feature vector may retain all numerical values while leaving their physiological relationships implicit, whereas a large end\-to\-end decoder may recover these relationships only through additional capacity and optimization\. We instead consider whether the representation itself can make the relevant structure directly accessible\. Under this view, a useful representation should satisfy two empirical criteria\. First, its advantage should remain visible under capacity\-controlled readouts rather than depending exclusively on model scaling\. Second, performance should deteriorate when the correspondence between temporal lag, channel identity, frequency semantics, and the fMRI target is deliberately disrupted\. These criteria turn the representation hypothesis into a directly testable claim\.

To test this hypothesis, we proposeMEL\(Multi\-bandEEGLatent\-state Tokenization\), a coordinate\-preserving EEG representation framework\. For each target fMRI response, MEL anchors the preceding EEG history, partitions it into temporal lag bins, computes channel\-wise canonical bandpower, applies train\-only calibration, and constructs lag\-channel\-frequency tokens\. The resulting representation can be decoded using Ridge regression or a compact multilayer perceptron\. MEL does not claim that EEG bandpower or lagged spectral analysis is itself new\. Instead, it organizes these classical measurements as an explicit target\-anchored coordinate system, rather than a temporally compressed summary or an encoding whose coordinate semantics are left implicit\.

MEL is expected to be particularly useful when prediction depends on distributed multi\-channel and multi\-band activity over an extended neural history\. Primary sensory ROIs may already be represented effectively by strong pretrained encoders because their responses can be comparatively localized or stimulus\-linked\. In contrast, high\-level cognitive, subcortical, and global targets may depend more strongly on distributed neural\-state dynamics across time\. We treat this anatomical pattern as an empirical expectation rather than a universal physiological assumption\.

Contributions\.Our main contributions are:

- •We formulate ROI\-level EEG\-to\-fMRI prediction as a*representation\-interface problem*, motivated by the delayed, distributed, and oscillatory nature of EEG\-fMRI coupling\.
- •We introduceMEL, a deterministic coordinate\-preserving tokenization framework that organizes pre\-target EEG histories into train\-calibrated lag\-channel\-frequency coordinates for downstream decoding\.
- •We evaluate MEL under intra\-scan, strict leave\-one\-subject\-out, and external\-dataset settings, together with capacity\-controlled readouts, correspondence\-breaking controls, and temporal and spectral ablations that characterize when and why the representation is effective\.

## Method

### Overview

MEL is designed as a representation interface rather than a high\-capacity predictor\. Given a target fMRI responseyty\_\{t\}, MEL anchors the preceding EEG segmentEt−H:tE\_\{t\-H:t\}, partitions it into lag bins, preserves channel identity, extracts canonical bandpower coordinates, applies train\-only calibration, and vectorizes the resulting token cube for capacity\-controlled ROI response decoding\. Bandpower is used as a stable physiological measurement; the methodological object is the task\-anchored coordinate system that preserves when the EEG evidence occurs, where it is recorded, and which oscillatory component it belongs to\.

![Refer to caption](https://arxiv.org/html/2608.29304v1/framework.png)Figure 2:Overall framework of MEL\. The method anchors each fMRI target to its preceding EEG history, constructs lag\-channel\-frequency tokens through train\-only calibrated bandpower extraction, and decodes the resulting representation into ROI responses using capacity\-controlled readouts\.
### Task Anchoring

LetE∈ℝT×CE\\in\\mathbb\{R\}^\{T\\times C\}denote an EEG recording withTTtime samples andCCchannels\. Letyt∈ℝRy\_\{t\}\\in\\mathbb\{R\}^\{R\}be the fMRI ROI response at target timett, whereRRis the number of target ROIs\. MEL forms each supervised sample by anchoring the target to a preceding EEG history window:

St=\(Et−H:t,ℛ,yt\),S\_\{t\}=\\left\(E\_\{t\-H:t\},\\mathcal\{R\},y\_\{t\}\\right\),\(1\)whereHHdenotes the EEG history length andℛ\\mathcal\{R\}is the ROI set\. This anchoring step is important because EEG\-to\-fMRI prediction is not a synchronous mapping\. It explicitly encodes the assumption thatyty\_\{t\}depends on a delayed EEG history\.

### Lag\-Channel Grid

The EEG historyEt−H:tE\_\{t\-H:t\}is divided into a set of temporal lag binsℒ=\{ℓ1,…,ℓL\}\\mathcal\{L\}=\\\{\\ell\_\{1\},\\ldots,\\ell\_\{L\}\\\}\. For each channelccand lag binℓ\\ell, MEL extracts a local EEG segment:

gc,ℓ=Et−ℓ−Δ:t−ℓ,c,g\_\{c,\\ell\}=E\_\{t\-\\ell\-\\Delta:t\-\\ell,c\},\(2\)whereΔ\\Deltais the bin duration\. The resulting lag\-channel grid is

Gt=\{gc,ℓ\}c=1,ℓ=1C,L∈ℝC×L\.G\_\{t\}=\\\{g\_\{c,\\ell\}\\\}\_\{c=1,\\ell=1\}^\{C,L\}\\in\\mathbb\{R\}^\{C\\times L\}\.\(3\)Unlike generic sequence patches, this grid preserves the identity of both the channel and the temporal delay relative to the fMRI target\.

### Bandpower Token Construction

For each lag\-channel segmentgc,ℓg\_\{c,\\ell\}, MEL computes bandpower over canonical EEG bandsℬ=\{δ,θ,α,β,γ\}\\mathcal\{B\}=\\\{\\delta,\\theta,\\alpha,\\beta,\\gamma\\\}\. Letϕf\\phi\_\{f\}denote the frequency support of bandff\. The band token is:

xc,ℓ,f=log⁡\(1\+1\|Iℓ\|​∑τ∈Iℓ\[ℬf​\(Ec\)​\(τ\)\]2\),x\_\{c,\\ell,f\}=\\log\\left\(1\+\\frac\{1\}\{\|I\_\{\\ell\}\|\}\\sum\_\{\\tau\\in I\_\{\\ell\}\}\\left\[\\mathcal\{B\}\_\{f\}\(E\_\{c\}\)\(\\tau\)\\right\]^\{2\}\\right\),\(4\)whereℬf​\(⋅\)\\mathcal\{B\}\_\{f\}\(\\cdot\)denotes the band\-pass filtering operator for theff\-th canonical EEG band,EcE\_\{c\}denotes the pre\-target EEG history of channelcc, andIℓI\_\{\\ell\}denotes the samples belonging to lag binℓ\\ell\.

The full MEL token cube is then:

Xt\\displaystyle X\_\{t\}=ΦMEL\(Gt\),\[Xt\]c,ℓ,f=xc,ℓ,f,\\displaystyle=\\Phi\_\{\\mathrm\{MEL\}\}\(G\_\{t\}\),\\qquad\[X\_\{t\}\]\_\{c,\\ell,f\}=x\_\{c,\\ell,f\},\(5\)Xt∈ℝC×L×F\.\\displaystyle X\_\{t\}\\in\\mathbb\{R\}^\{C\\times L\\times F\}\.HereF=\|ℬ\|F=\|\\mathcal\{B\}\|, and each elementxc,ℓ,fx\_\{c,\\ell,f\}is a coordinate\-aware neural\-state token indexed by one EEG channel, one hemodynamic lag bin, and one frequency band\.

Algorithm 1MEL Tokenization and ROI Response Decoding0:EEG recording

EE, target time

tt, ROI set

ℛ\\mathcal\{R\}, history length

HH, lag bins

ℒ\\mathcal\{L\}, frequency bands

ℬ\\mathcal\{B\}, train statistics

μtr,σtr\\mu^\{\\mathrm\{tr\}\},\\sigma^\{\\mathrm\{tr\}\}
0:Predicted ROI response

𝐲^t\\hat\{\\mathbf\{y\}\}\_\{t\}
1:Anchor target sample

St=\(Et−H:t,ℛ,yt\)S\_\{t\}=\(E\_\{t\-H:t\},\\mathcal\{R\},y\_\{t\}\)
2:Partition

Et−H:tE\_\{t\-H:t\}into lag\-channel segments

\{gc,ℓ\}c=1,ℓ=1C,L\\\{g\_\{c,\\ell\}\\\}\_\{c=1,\\ell=1\}^\{C,L\}
3:Initialize calibrated token cube

X~t∈ℝC×L×F\\tilde\{X\}\_\{t\}\\in\\mathbb\{R\}^\{C\\times L\\times F\}
4:foreach coordinate triple

\(c,ℓ,f\)∈\{1,…,C\}×ℒ×ℬ\(c,\\ell,f\)\\in\\\{1,\\ldots,C\\\}\\times\\mathcal\{L\}\\times\\mathcal\{B\}do

5:Extract segment

gc,ℓg\_\{c,\\ell\}and apply band\-pass operator

ℬf\\mathcal\{B\}\_\{f\}
6:Compute log\-bandpower token

xc,ℓ,fx\_\{c,\\ell,f\}from the bandpower definition

7:Calibrate token

x~c,ℓ,f=\(xc,ℓ,f−μc,ℓ,ftr\)/\(σc,ℓ,ftr\+ϵ\)\\tilde\{x\}\_\{c,\\ell,f\}=\(x\_\{c,\\ell,f\}\-\\mu^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\)/\(\\sigma^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\+\\epsilon\)
8:Store

\[X~t\]c,ℓ,f←x~c,ℓ,f\[\\tilde\{X\}\_\{t\}\]\_\{c,\\ell,f\}\\leftarrow\\tilde\{x\}\_\{c,\\ell,f\}
9:endfor

10:Vectorize tokens with fixed order

xtflat=vec⁡\(X~t\)x\_\{t\}^\{\\mathrm\{flat\}\}=\\mathrm\{vec\}\(\\tilde\{X\}\_\{t\}\)
11:Decode ROI response

𝐲^t=gθ​\(xtflat\)\\hat\{\\mathbf\{y\}\}\_\{t\}=g\_\{\\theta\}\(x\_\{t\}^\{\\mathrm\{flat\}\}\)
12:return

𝐲^t\\hat\{\\mathbf\{y\}\}\_\{t\}

### Train\-only Calibration

To avoid leakage, normalization statistics are estimated only from the training split\. Letμc,ℓ,ftr\\mu^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}andσc,ℓ,ftr\\sigma^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}denote the train\-set mean and standard deviation for each MEL coordinate\. The normalized token is:

x~c,ℓ,f=xc,ℓ,f−μc,ℓ,ftrσc,ℓ,ftr\+ϵ\.\\tilde\{x\}\_\{c,\\ell,f\}=\\frac\{x\_\{c,\\ell,f\}\-\\mu^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\}\{\\sigma^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\+\\epsilon\}\.\(6\)This produces the calibrated token cubeX~t\\tilde\{X\}\_\{t\}\. All validation, test, and external samples are transformed using the same train\-set statistics\.

### ROI Response Decoding

MEL is model\-agnostic\. After calibration, the token cube is vectorized in a fixed coordinate order:

xtflat=vec⁡\(X~t\)∈ℝC​L​F\.x\_\{t\}^\{\\mathrm\{flat\}\}=\\mathrm\{vec\}\(\\tilde\{X\}\_\{t\}\)\\in\\mathbb\{R\}^\{CLF\}\.\(7\)The predicted ROI response is:

y^t=gθ​\(xtflat\)\.\\hat\{y\}\_\{t\}=g\_\{\\theta\}\(x\_\{t\}^\{\\mathrm\{flat\}\}\)\.\(8\)We evaluate Ridge regression and compact MLP readouts\. Ridge regression solves:

min⁡∑tW,b⁡‖yt−W​xtflat−b‖22\+λ​‖W‖22,\\min\_\{W,b\}\\sum\_\{t\}\\left\\\|y\_\{t\}\-Wx\_\{t\}^\{\\mathrm\{flat\}\}\-b\\right\\\|\_\{2\}^\{2\}\+\\lambda\\\|W\\\|\_\{2\}^\{2\},\(9\)while the MLP readout uses a compact nonlinear mapping\. Since the readouts are capacity\-controlled, strong performance indicates that MEL exposes predictive structure before the decoder rather than merely relying on large model capacity\.

### Algorithmic Properties

MEL is intentionally designed as a deterministic representation algorithm\. Algorithm 1 outlines the detailed steps required to execute this tokenization procedure\. Rather than introducing a new optimization objective, it changes the coordinate system in which EEG history is presented to the decoder\. This gives two useful properties for EEG\-to\-fMRI prediction and clarifies why the algorithm is stable across readout choices\.

Proposition 1 \(Coordinate preservation\)\.Given a fixed lag setℒ\\mathcal\{L\}, channel set𝒞\\mathcal\{C\}, band bankℬ\\mathcal\{B\}, and vectorization order, MEL defines a deterministic mapping

ΦMEL:Et−H:t↦X~t∈ℝC×L×F\.\\Phi\_\{\\mathrm\{MEL\}\}:E\_\{t\-H:t\}\\mapsto\\tilde\{X\}\_\{t\}\\in\\mathbb\{R\}^\{C\\times L\\times F\}\.\(10\)Each tokenx~c,ℓ,f\\tilde\{x\}\_\{c,\\ell,f\}has a unique semantic coordinate corresponding to one channel, one temporal lag, and one frequency band\. Therefore, vectorization does not destroy coordinate identity; it only changes storage order\.

*Proof sketch\.*Each step of MEL is indexed by an explicit coordinate: lag extraction is indexed byℓ\\ell, channel selection bycc, and bandpower extraction byff\. With a fixed vectorization order,vec⁡\(X~t\)\\mathrm\{vec\}\(\\tilde\{X\}\_\{t\}\)is a bijective re\-indexing of the token cube rather than a lossy pooling operation\. Thus, coordinate identity is preserved until the readout stage\.

*Implication\.*This property makes MEL different from generic EEG embeddings whose internal dimensions are not directly interpretable\. It allows destructive controls such as lag reversal, band reversal, and channel permutation to test specific coordinates of the representation\.

Proposition 2 \(Split\-stable calibration\)\.If calibration statisticsμtr\\mu^\{\\mathrm\{tr\}\}andσtr\\sigma^\{\\mathrm\{tr\}\}are estimated only from the training split, then the MEL transform applied to validation, test, or external samples does not use target\-split statistics:

x~c,ℓ,feval=xc,ℓ,feval−μc,ℓ,ftrσc,ℓ,ftr\+ϵ\.\\tilde\{x\}\_\{c,\\ell,f\}^\{\\mathrm\{eval\}\}=\\frac\{x\_\{c,\\ell,f\}^\{\\mathrm\{eval\}\}\-\\mu^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\}\{\\sigma^\{\\mathrm\{tr\}\}\_\{c,\\ell,f\}\+\\epsilon\}\.\(11\)Thus, the representation is deterministic after training calibration and does not introduce normalization leakage\.

*Proof sketch\.*For any evaluation sample,x~c,ℓ,feval\\tilde\{x\}\_\{c,\\ell,f\}^\{\\mathrm\{eval\}\}depends only on its own unnormalized token and onμtr,σtr\\mu^\{\\mathrm\{tr\}\},\\sigma^\{\\mathrm\{tr\}\}\. No statistic computed from validation, test, or external targets enters the transform\. Therefore, the calibration operator is split\-stable once the training split is fixed\.

*Implication\.*This property supports reproducibility and makes MEL suitable for capacity\-controlled evaluation\. If Ridge regression or a compact MLP improves prediction on train\-calibrated MEL tokens, the gain is more plausibly attributed to representation structure than to test\-set normalization or decoder scaling\.

Evaluation principle \(Capacity\-controlled readability\)\.MEL is evaluated with a hierarchy of readouts: Ridge regression tests whether the token space is linearly readable, a compact MLP tests whether mild nonlinear mixing is sufficient, and an ensemble tests stability under small readout variations\. This hierarchy is not intended to search for a larger architecture\. Instead, it tests whether the representation itself exposes useful EEG\-fMRI structure before the decoder\.

## Experiments

We evaluate MEL on the VU EEG\-fMRI benchmark and external Oddball and NODDI settings through four research questions:

RQ1\.Does MEL improve EEG\-to\-fMRI prediction over neural and traditional representation baselines? RQ2\.Does the representation remain useful under subject and external\-dataset shifts? RQ3\.Does the gain depend on valid EEG\-fMRI correspondence rather than leakage or shortcut statistics? RQ4\.Which temporal, spectral, and anatomical components account for the improvement?

### Experimental Setup

Datasets\.Our primary benchmark is the VU simultaneous EEG\-fMRI dataset used by NeuroBOLT\([Li et al\. 2024b](https://arxiv.org/html/2608.29304#bib.bib31)\), containing 29 scans from 22 subjects\. For each fMRI time point, the preceding 16 seconds of EEG are used to predict seven targets: cuneus, Heschl’s gyrus, anterior middle frontal gyrus, anterior precuneus, putamen, thalamus, and the global signal\. We further evaluate external transfer on the Auditory and Visual Oddball dataset\([Walz et al\. 2018](https://arxiv.org/html/2608.29304#bib.bib43)\)and use NODDI\([EBRAINS 2024](https://arxiv.org/html/2608.29304#bib.bib14)\)as a stricter alignment\-quality stress test\.

Protocols and metrics\.We follow the NeuroBOLT intra\-scan and inter\-subject protocols\. Intra\-scan evaluation uses predefined train, validation, and test partitions within each scan\. Inter\-subject evaluation follows strict leave\-one\-subject\-out testing, where all scans of the held\-out subject are excluded from training and validation\. Pearson correlationRRis computed for each target, and Avg\.RRdenotes the mean over the seven targets\. Normalization statistics, checkpoint selection, Ridge penalties, and readout settings are determined from training and validation data only; test data are never used for normalization, model selection, or seed selection\.

Table 1:EEG\-to\-fMRI prediction under intra\-scan and inter\-subject evaluation\. Entries report mean±\\pmstandard deviation, and Avg\.RRis averaged over the seven targets\. Red and blue denote the best and second\-best results within each setting\. All neural and traditional baselines are reproduced under the same preprocessing, split, train\-only normalization, and validation\-based model\-selection pipeline rather than taken from incompatible reporting protocols\. BP denotes bandpower, and inter\-subject results follow strict leave\-one\-subject\-out evaluation over 22 subjects and 29 scans\.ModelPrimary SensoryHigh\-level CognitiveSubcortical–Avg\. R↑\\uparrowCuneusHeschl’s GyrusMiddle FrontalPrecuneus AnteriorPutamenThalamusGlobal SignalMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowMSE↓\\downarrowR↑\\uparrowT2T\_\{2\}0\.4240\.1640\.3530\.1650\.3240\.3950\.3530\.3580\.4460\.2420\.3510\.3330\.3560\.3830\.292T1T\_\{1\}0\.3250\.2210\.2990\.1820\.3140\.4470\.3020\.3900\.3100\.2520\.3230\.3730\.2810\.4100\.325M\+B3M\+B\_\{3\}0\.3150\.3830\.2910\.3050\.3140\.4510\.3220\.4660\.3410\.3430\.3450\.4100\.3020\.4890\.407M\+B5M\+B\_\{5\}0\.2590\.4340\.2630\.3330\.2650\.5160\.2480\.5020\.3200\.3730\.2770\.4550\.2470\.5510\.452

Table 2:Ablation of MEL components\.MMdenotes the lag\-channel\-band coordinate system\.B3B\_\{3\}andB5B\_\{5\}use three and five frequency bands, respectively\.T1T\_\{1\}andT2T\_\{2\}denote HRF\-weighted and time\-mean aggregation\. All variants share the same Ridge readout\.Baselines\.Neural baselines include BIOT\([Yang et al\. 2023](https://arxiv.org/html/2608.29304#bib.bib47)\), LaBraM\([Jiang et al\. 2024](https://arxiv.org/html/2608.29304#bib.bib23)\), BEIRA\([Kovalev et al\. 2022](https://arxiv.org/html/2608.29304#bib.bib26)\), SIREN\([Li et al\. 2024a](https://arxiv.org/html/2608.29304#bib.bib30)\), and NeuroBOLT\([Li et al\. 2024b](https://arxiv.org/html/2608.29304#bib.bib31)\)\. The inter\-subject benchmark additionally includes FFCL, CNN Transformer, and STT Transformer\([Li et al\. 2022](https://arxiv.org/html/2608.29304#bib.bib29);[Peh et al\. 2022](https://arxiv.org/html/2608.29304#bib.bib36);[Song et al\. 2021](https://arxiv.org/html/2608.29304#bib.bib42)\)\. We also implement five traditional spectral representations with the same Ridge readout: classical time\-mean bandpower, HRF\-weighted bandpower, flattened power spectral density, classical lagged bandpower, and uniform STFT lag\-band features\. All results in Table[1](https://arxiv.org/html/2608.29304#Sx3.T1)are reproduced or re\-evaluated under identical preprocessing, splits, train\-only normalization, and validation\-based model selection\.

Implementation\.The default MEL representation uses a 16\-second pre\-target EEG history, 2\-second lag bins, 26 channels, and five canonical EEG bands\. Ridge penalties are selected independently for each target from 31 logarithmically spaced candidates using validation data\. Compact MLP readouts are trained with seeds\{1,2,3,4,5\}\\\{1,2,3,4,5\\\}, and MEL\-MLP Ensemble averages their predictions\. Experiments were executed in the same software environment on an NVIDIA GeForce RTX 4060 Laptop GPU with 8 GB memory; closed\-form Ridge fitting does not require GPU acceleration\.

Figure 3:External transfer at different temporal depths\. Oddball performance improves as a longer ordered EEG history is retained, whereas NODDI remains near zero under the available alignment cache\.Table 3:Representation\-validity controls\. Each experiment keeps the Ridge readout fixed while disrupting one component of MEL\.Figure 4:Temporal\-design ablation\. Bars report test Avg\.RR, while black curves report validation Avg\.RR\.![Refer to caption](https://arxiv.org/html/2608.29304v1/mel_otkd_style_coordinate_sensitivity.png)Figure 5:MEL coordinate\-sensitivity summary across frequency bands, temporal lags, and ROI groups\. Values combine normalized marginal band\-ablation, lag\-ablation, and ROI\-gain diagnostics within the displayed map\.
### RQ1: Comparison Experiments

Under intra\-scan evaluation, MEL\-MLP Ensemble achieves the highest Avg\.RRof 0\.543, improving the NeuroBOLT reference from 0\.531\. The largest gains appear on middle frontal, anterior precuneus, putamen, thalamus, and global signal, while NeuroBOLT remains strongest on cuneus and Heschl’s gyrus\. The improvement is therefore structured rather than uniform across ROIs\.

The traditional controls clarify why this gain is nontrivial\. Classical time\-mean and HRF\-weighted bandpower reach only 0\.258 and 0\.312 Avg\.RR, and the strongest conventional lag\-resolved spectral baseline, uniform STFT lag\-band, reaches 0\.455\. MEL\-Ridge remains competitive at 0\.452, while MEL\-MLP and MEL\-MLP Ensemble raise Avg\.RRto 0\.524 and 0\.543\. Thus, the improvement is not explained by bandpower alone or by an oversized decoder; it appears when lag, channel, and band coordinates are preserved as an explicit decoding interface\.

Figure 6:Examples of fMRI ROI reconstruction on unseen scans\. Dashed gray curves denote ground\-truth responses, and colored curves denote MEL predictions\.We further run a matched\-readout audit in which classical lagged bandpower, uniform STFT lag\-band features, flattened PSD features, and MEL\-coordinate tokens use the same MLP protocol and train\-only calibration\. Under this capacity\-controlled comparison, MEL\-coordinate tokens retain the strongest performance, whereas matched classical lagged bandpower and STFT controls drop substantially\. This audit directly addresses whether MEL is merely a renaming of conventional bandpower regression: the useful signal comes from the task\-aligned coordinate interface rather than from bandpower measurement alone\.

### RQ2: Generalization Experiments

Under inter\-subject evaluation, MEL\-MLP Ensemble reaches an Avg\.RRof 0\.499, improving on NeuroBOLT’s 0\.473\. The target\-wise pattern remains consistent with intra\-scan evaluation: MEL is strongest on the high\-level cognitive, subcortical, and global targets, while pretrained neural encoders remain strongest on the two sensory ROIs\. This consistency argues against a uniform correlation\-inflation effect\.

The traditional inter\-subject baselines remain substantially lower under the matched leave\-one\-subject\-out protocol\. Uniform STFT lag\-band is the strongest conventional representation at 0\.382 Avg\.RR, followed by classical lagged bandpower at 0\.377\. MEL\-Ridge reaches 0\.439 and outperforms all five traditional representations across all seven targets, while MEL\-MLP and its ensemble further improve Avg\.RRto 0\.488 and 0\.499\. Thus, the representation remains useful even when the target subject is unseen during training\.

Figure[3](https://arxiv.org/html/2608.29304#Sx3.F3)extends the evaluation beyond VU\. Oddball performance improves as the representation retains a longer ordered EEG history, supporting lag\-resolved tokenization under a different task and recording configuration\. NODDI remains near zero across temporal depths\. We therefore draw a bounded conclusion: MEL transfers when a usable EEG\-fMRI alignment substrate is present, but tokenization cannot recover structure from severely degraded external correspondence\.

### RQ3: Representation Validity

Prediction gains alone do not establish that MEL uses the intended coordinates\. We therefore keep the Ridge readout fixed while selectively breaking normalization, label correspondence, temporal order, channel identity, and spectral semantics\.

Table[3](https://arxiv.org/html/2608.29304#Sx3.T3)provides a direct answer\. Train\-only and all\-split normalization produce the same score \(0\.45210\.4521\), excluding normalization leakage as the source of improvement\. Label and target shuffling collapse Avg\.RRto0\.00140\.0014and−0\.0021\-0\.0021, showing that prediction requires genuine EEG\-fMRI correspondence\. Channel permutation reduces Avg\.RRto0\.39450\.3945, band reversal to0\.23550\.2355, random lag assignment to0\.19770\.1977, and lag reversal to0\.10220\.1022\. This ordering is the key validity result: innocuous preprocessing changes do not help, whereas breaking the intended coordinates destroys performance\.

### RQ4: Design Attribution and Mechanism

Component attribution\.Table[2](https://arxiv.org/html/2608.29304#Sx3.T2)compares increasingly structured alternatives under the same Ridge decoder\. Time\-mean and HRF\-weighted aggregation reach Avg\.RRvalues of 0\.292 and 0\.325\. Preserving lag coordinates with three bands raises performance to 0\.407, while the full five\-band representation reaches 0\.452\. The ordering

T2<T1<M\+B3<M\+B5T\_\{2\}<T\_\{1\}<M\+B\_\{3\}<M\+B\_\{5\}shows that MEL benefits from preserving both ordered temporal support and adequate spectral coverage\.

Temporal design\.Figure[4](https://arxiv.org/html/2608.29304#Sx3.F4)shows that 2\-second bins provide the best balance between temporal resolution and stable power estimation\. Shorter bins produce noisier estimates, whereas longer bins merge distinct parts of the EEG history\. The full pre\-target window performs best, while 5–11 s is the strongest restricted range\. This pattern supports a distributed temporal dependency rather than a single fixed delay, and it explains why the full lag\-resolved representation outperforms a predetermined HRF\-weighted summary\.

Frequency\-lag sensitivity\.We define sensitivity using held\-out feature ablation rather than gradients\. For a band or lag subsetqq, the trained readout is kept fixed and the corresponding MEL coordinates are replaced by their train\-set mean; the score is the scan\-averaged drop in test Avg\.RR\. ROI weights are computed from ROI\-wise prediction gains, and Figure[5](https://arxiv.org/html/2608.29304#Sx3.F5)visualizes the normalized product of these marginal band, lag, and ROI diagnostics\. The map therefore summarizes where independent evidence co\-concentrates in the token cube, rather than claiming a separately trained three\-way interaction model\. The strongest sensitivity appears in the beta band and around the 7–9 s lag range, consistent with the validity controls where disrupting lag order and band semantics causes the largest performance losses\.

Anatomical behavior\.Figure[6](https://arxiv.org/html/2608.29304#Sx3.F6)shows that MEL captures ROI\-specific dynamics rather than a shared global trend, with its strongest advantages concentrated in cognitive, subcortical, and global targets, while pretrained encoders remain stronger on sensory ROIs\.

Taken together, these results support a specific interpretation of novelty: MEL is not simply a generic bandpower feature stack\. Its advantage depends on preserving lag, band, and channel semantics as an explicit coordinate system, and that advantage remains visible under subject shift and controlled external transfer\.

## Conclusion

We presented MEL, a coordinate\-preserving EEG tokenization framework for EEG\-to\-fMRI translation\. MEL organizes pre\-target EEG histories into explicit lag\-channel\-frequency coordinates, reducing the representation mismatch between fast electrophysiological activity and delayed hemodynamic responses\. Across intra\-scan and inter\-subject evaluation, MEL outperforms strong neural and traditional baselines\. Ridge results show that the representation is directly readable, while lightweight nonlinear readouts capture additional cross\-coordinate interactions\. Oddball transfer, ablations, and validity controls further demonstrate that the gains depend on ordered temporal, spectral, and spatial structure rather than leakage or decoder capacity\. These results establish representation design as an important direction for EEG\-to\-fMRI research\. Although MEL does not eliminate cross\-subject variability or imperfect multimodal alignment, it provides a simple, interpretable, and model\-compatible interface for exposing fMRI\-relevant EEG structure\.

## References

- Abreuet al\.\(2018\)R\. Abreu, A\. Leal, and P\. FigueiredoEEG\-informed fMRI: a review of data analysis methods\.Frontiers in Human Neuroscience12,pp\. 29\.External Links:[Document](https://dx.doi.org/10.3389/fnhum.2018.00029)
- Afrasiyabiet al\.\(2025\)A\. Afrasiyabi, D\. Bhaskar, E\. L\. Busch, L\. Caplette, R\. Singh, G\. Lajoie, N\. B\. Turk\-Browne, and S\. KrishnaswamyLatent representation learning for multimodal brain activity translation\.InICASSP 2025 – 2025 IEEE International Conference on Acoustics, Speech and Signal Processing,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/ICASSP49660.2025.10887834)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Biswalet al\.\(1995\)B\. Biswal, F\. Z\. Yetkin, V\. M\. Haughton, and J\. S\. HydeFunctional connectivity in the motor cortex of resting human brain using echo\-planar MRI\.Magnetic Resonance in Medicine34\(4\),pp\. 537–541\.External Links:[Document](https://dx.doi.org/10.1002/mrm.1910340409)
- Buxtonet al\.\(1998\)R\. B\. Buxton, E\. C\. Wong, and L\. R\. FrankDynamics of blood flow and oxygenation changes during brain activation: the balloon model\.Magnetic Resonance in Medicine39\(6\),pp\. 855–864\.External Links:[Document](https://dx.doi.org/10.1002/mrm.1910390602)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Buzsáki and Draguhn \(2004\)G\. Buzsáki and A\. DraguhnNeuronal oscillations in cortical networks\.Science304\(5679\),pp\. 1926–1929\.External Links:[Document](https://dx.doi.org/10.1126/science.1099745)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Calhas and Henriques \(2020\)D\. Calhas and R\. HenriquesEEG to fMRI synthesis: is deep learning a candidate?\.InInformation Systems Development, Organizational Aspects and Societal Trends \(ISD2023 Proceedings\),External Links:[Document](https://dx.doi.org/10.62036/ISD.2023.26)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Calhas and Henriques \(2022\)D\. Calhas and R\. HenriquesEEG to fMRI synthesis benefits from attentional graphs of electrode relationships\.arXiv preprint arXiv:2203\.03481\.External Links:[Link](https://proceedings.mlr.press/v219/calhas23a.html)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Chenet al\.\(2026\)W\. Chen, Y\. Wu, J\. Chen, X\. Zhou, and Y\. LiangUrbanFM: scaling urban spatio\-temporal foundation models\.arXiv preprint arXiv:2602\.20677\.
- Cohen \(2014\)M\. X\. CohenAnalyzing neural time series data: theory and practice\.MIT Press,Cambridge, MA\.External Links:ISBN 9780262019873,[Link](https://mitpress.mit.edu/9780262019873/analyzing-neural-time-series-data/)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Damoiseauxet al\.\(2006\)J\. S\. Damoiseaux, S\. A\. R\. B\. Rombouts, F\. Barkhof, P\. Scheltens, C\. J\. Stam, S\. M\. Smith, and C\. F\. BeckmannConsistent resting\-state networks across healthy subjects\.Proceedings of the National Academy of Sciences103\(37\),pp\. 13848–13853\.External Links:[Document](https://dx.doi.org/10.1073/pnas.0601417103)
- de Muncket al\.\(2009\)J\. C\. de Munck, S\. I\. Gonçalves, R\. Mammoliti, R\. M\. Heethaar, and F\. L\. Da SilvaInteractions between different eeg frequency bands and their effect on alpha–fmri correlations\.Neuroimage47\(1\),pp\. 69–76\.
- de Muncket al\.\(2008\)J\. C\. de Munck, S\. I\. Gonçalves, T\. J\. C\. Faes, J\. P\. A\. Kuijer, P\. J\. W\. Pouwels, R\. M\. Heethaar, and F\. H\. Lopes da SilvaA study of the brain’s resting state based on alpha band power, heart rate and fMRI\.NeuroImage42\(1\),pp\. 112–121\.External Links:[Document](https://dx.doi.org/10.1016/j.neuroimage.2008.04.244)
- Debeneret al\.\(2005\)S\. Debener, M\. Ullsperger, M\. Siegel, K\. Fiehler, D\. Y\. von Cramon, and A\. K\. EngelTrial\-by\-trial coupling of concurrent electroencephalogram and functional magnetic resonance imaging identifies the dynamics of performance monitoring\.Journal of Neuroscience25\(50\),pp\. 11730–11737\.External Links:[Document](https://dx.doi.org/10.1523/JNEUROSCI.3286-05.2005)
- Donoso \(2025\)M\. DonosoNeural networks and foundation models: two strategies for EEG\-to\-fMRI prediction\.Frontiers in Systems Biology5,pp\. 1715692\.External Links:[Document](https://dx.doi.org/10.3389/fsysb.2025.1715692)
- EBRAINS \(2024\)EBRAINSEEG, fMRI and NODDI at rest\.Note:EBRAINS Knowledge Graph DatasetAccessed: 2026\-06\-25External Links:[Link](https://search.kg.ebrains.eu/instances/5706c36d-ec44-45bd-a2c7-230b5cc5f557)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p1.1)\.
- Eicheleet al\.\(2005\)T\. Eichele, K\. Specht, M\. Moosmann, M\. L\. A\. Jongsma, R\. Q\. Quiroga, H\. Nordby, and K\. HugdahlAssessing the spatiotemporal evolution of neuronal activation with single\-trial event\-related potentials and functional MRI\.Proceedings of the National Academy of Sciences102\(49\),pp\. 17798–17803\.External Links:[Document](https://dx.doi.org/10.1073/pnas.0505508102)
- Engel and Fries \(2010\)A\. K\. Engel and P\. FriesBeta\-band oscillations: signalling the status quo?\.Current Opinion in Neurobiology20\(2\),pp\. 156–165\.External Links:[Document](https://dx.doi.org/10.1016/j.conb.2010.02.015)
- Fox and Raichle \(2007\)M\. D\. Fox and M\. E\. RaichleSpontaneous fluctuations in brain activity observed with functional magnetic resonance imaging\.Nature Reviews Neuroscience8\(9\),pp\. 700–711\.External Links:[Document](https://dx.doi.org/10.1038/nrn2201)
- Glover \(1999\)G\. H\. GloverDeconvolution of impulse response in event\-related BOLD fMRI\.NeuroImage9\(4\),pp\. 416–429\.External Links:[Document](https://dx.doi.org/10.1006/nimg.1998.0419)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Goldmanet al\.\(2002\)R\. I\. Goldman, J\. M\. Stern, J\. Jr\. Engel, and M\. S\. CohenSimultaneous EEG and fMRI of the alpha rhythm\.NeuroReport13\(18\),pp\. 2487–2492\.External Links:[Document](https://dx.doi.org/10.1097/00001756-200212200-00022)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Heet al\.\(2025\)D\. He, S\. Li, B\. Jiang, and H\. YanSpec2VolCAMU\-Net: a spectrogram\-to\-volume model for EEG\-to\-fMRI reconstruction based on multi\-directional time\-frequency convolutional attention encoder and vision\-mamba U\-net\.Journal of Neural Engineering22\(5\),pp\. 056042\.External Links:[Link](https://arxiv.org/abs/2505.09521)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Hoerl and Kennard \(2000\)A\. E\. Hoerl and R\. W\. KennardRidge regression: biased estimation for nonorthogonal problems\.Technometrics42\(1\),pp\. 80–86\.External Links:[Document](https://dx.doi.org/10.1080/00401706.1970.10488634)
- Husteret al\.\(2012\)R\. J\. Huster, S\. Debener, T\. Eichele, and C\. S\. HerrmannMethods for simultaneous EEG\-fMRI: an introductory review\.Journal of Neuroscience32\(18\),pp\. 6053–6060\.External Links:[Document](https://dx.doi.org/10.1523/JNEUROSCI.0447-12.2012)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Jianget al\.\(2024\)W\. Jiang, L\. Zhao, and B\. LuLarge brain model for learning generic representations with tremendous EEG data in BCI\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 16405–16426\.External Links:[Link](https://openreview.net/forum?id=QzTpTRVtrP)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.20.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.4.1)\.
- Jorgeet al\.\(2014\)J\. Jorge, W\. Van der Zwaag, and P\. FigueiredoEEG\-fMRI integration for the study of human brain function\.NeuroImage102,pp\. 24–34\.External Links:[Document](https://dx.doi.org/10.1016/j.neuroimage.2013.05.114)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Klimesch \(1999\)W\. KlimeschEEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis\.Brain Research Reviews29\(2–3\),pp\. 169–195\.External Links:[Document](https://dx.doi.org/10.1016/S0165-0173%2898%2900056-3)
- Kovalevet al\.\(2022\)A\. Kovalev, I\. Mikheev, and A\. OssadtchifMRI from EEG is only deep learning away: the use of interpretable DL to unravel EEG\-fMRI relationships\.arXiv preprint arXiv:2211\.02024\.External Links:[Link](https://arxiv.org/abs/2211.02024)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.21.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.5.1)\.
- Lanzinoet al\.\(2024\)R\. Lanzino, F\. Fontana, L\. Cinque, F\. Scarcello, and A\. MakiNT\-ViT: neural transcoding vision transformers for EEG\-to\-fMRI synthesis\.arXiv preprint arXiv:2409\.11836\.External Links:[Link](https://arxiv.org/abs/2409.11836)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Laufset al\.\(2003\)H\. Laufs, A\. Kleinschmidt, A\. Beyerle, E\. Eger, A\. Salek\-Haddadi, C\. Preibisch, and K\. KrakowEEG\-correlated fMRI of human alpha activity\.NeuroImage19\(4\),pp\. 1463–1476\.External Links:[Document](https://dx.doi.org/10.1016/S1053-8119%2803%2900286-6)
- Liet al\.\(2022\)H\. Li, M\. Ding, R\. Zhang, and C\. XiuMotor imagery EEG classification algorithm based on CNN\-LSTM feature fusion network\.Biomedical Signal Processing and Control72,pp\. 103342\.External Links:[Document](https://dx.doi.org/10.1016/j.bspc.2021.103342)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.16.2)\.
- Liet al\.\(2024a\)Y\. Li, A\. Lou, Z\. Xu, S\. Wang, and C\. ChangLeveraging sinusoidal representation networks to predict fMRI signals from EEG\.InMedical Imaging 2024: Image Processing,Vol\.12926,pp\. 795–800\.Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.22.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.6.1)\.
- Liet al\.\(2024b\)Y\. Li, A\. Lou, Z\. Xu, S\. Zhang, S\. Wang, D\. J\. Englot, S\. Kolouri, D\. Moyer, R\. G\. Bayrak, and C\. ChangNeuroBOLT: resting\-state EEG\-to\-fMRI synthesis with multi\-dimensional feature mapping\.arXiv preprint arXiv:2410\.05341\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2024/hash/29c294ddc628c94cd2c636383ef106c1-Abstract-Conference.html)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1),[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p1.1),[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.23.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.7.1)\.
- Liu and Sajda \(2023\)X\. Liu and P\. SajdaLatent neural source recovery via transcoding of simultaneous EEG\-fMRI\.InInternational Conference on Brain Informatics,pp\. 318–330\.External Links:[Document](https://dx.doi.org/10.1007/978-3-031-43075-6%5F28)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Logothetiset al\.\(2001\)N\. K\. Logothetis, J\. Pauls, M\. Augath, T\. Trinath, and A\. OeltermannNeurophysiological investigation of the basis of the fMRI signal\.Nature412\(6843\),pp\. 150–157\.External Links:[Document](https://dx.doi.org/10.1038/35084005)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Mantiniet al\.\(2007\)D\. Mantini, M\. G\. Perrucci, C\. Del Gratta, G\. L\. Romani, and M\. CorbettaElectrophysiological signatures of resting state networks in the human brain\.Proceedings of the National Academy of Sciences104\(32\),pp\. 13170–13175\.External Links:[Document](https://dx.doi.org/10.1073/pnas.0700668104)
- Moosmannet al\.\(2003\)M\. Moosmann, P\. Ritter, I\. Krastel, A\. Brink, S\. Thees, F\. Blankenburg, B\. Taskin, H\. Obrig, and A\. VillringerCorrelates of alpha rhythm in functional magnetic resonance imaging and near infrared spectroscopy\.NeuroImage20\(1\),pp\. 145–158\.External Links:[Document](https://dx.doi.org/10.1016/S1053-8119%2803%2900344-6)
- Pehet al\.\(2022\)W\. Y\. Peh, Y\. Yao, and J\. DauwelsTransformer convolutional neural networks for automated artifact detection in scalp EEG\.In2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society,pp\. 3599–3602\.External Links:[Document](https://dx.doi.org/10.1109/EMBC48229.2022.9871916)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.17.1)\.
- Pfurtscheller and Lopes da Silva \(1999\)G\. Pfurtscheller and F\. H\. Lopes da SilvaEvent\-related EEG/MEG synchronization and desynchronization: basic principles\.Clinical Neurophysiology110\(11\),pp\. 1842–1857\.External Links:[Document](https://dx.doi.org/10.1016/S1388-2457%2899%2900141-8)
- Philiastideset al\.\(2021\)M\. G\. Philiastides, T\. Tu, and P\. SajdaInferring macroscale brain dynamics via fusion of simultaneous EEG\-fMRI\.Annual Review of Neuroscience44\(1\),pp\. 315–334\.External Links:[Document](https://dx.doi.org/10.1146/annurev-neuro-100220-093239)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Rooset al\.\(2025\)K\. Roos, A\. Fukuda, and Q\. H\. CapFrom brainwaves to brain scans: a robust neural network for EEG\-to\-fMRI synthesis\.arXiv preprint arXiv:2502\.08025\.External Links:[Document](https://dx.doi.org/10.1109/EMBC58623.2025.11253321)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.
- Rosaet al\.\(2010\)M\. J\. Rosa, J\. Daunizeau, and K\. J\. FristonEEG\-fMRI integration: a critical review of biophysical modeling and data analysis approaches\.Journal of Integrative Neuroscience9\(04\),pp\. 453–476\.External Links:[Document](https://dx.doi.org/10.1142/S0219635210002512)
- Scheeringaet al\.\(2011\)R\. Scheeringa, P\. Fries, K\. Petersson, R\. Oostenveld, I\. Grothe, D\. G\. Norris, P\. Hagoort, and M\. C\. M\. BastiaansenNeuronal dynamics underlying high\- and low\-frequency EEG oscillations contribute independently to the human BOLD signal\.Neuron69\(3\),pp\. 572–583\.External Links:[Document](https://dx.doi.org/10.1016/j.neuron.2010.11.044)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p1.1)\.
- Songet al\.\(2021\)Y\. Song, X\. Jia, L\. Yang, and L\. XieTransformer\-based spatial\-temporal feature learning for EEG decoding\.arXiv preprint arXiv:2106\.11170\.External Links:[Link](https://arxiv.org/abs/2106.11170)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.18.1)\.
- Walzet al\.\(2018\)J\. M\. Walz, R\. I\. Goldman, J\. Muraskin, B\. Conroy, T\. R\. Brown, and P\. SajdaAuditory and visual oddball EEG\-fMRI\.Note:OpenNeuro, ds000116Accessed: 2026\-06\-25External Links:[Link](https://openneuro.org/datasets/ds000116)Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p1.1)\.
- Warbrick \(2022\)T\. WarbrickSimultaneous EEG\-fMRI: what have we learned and what does the future hold?\.Sensors22\(6\),pp\. 2262\.External Links:[Document](https://dx.doi.org/10.3390/s22062262)
- Weiet al\.\(2025\)X\. Wei, K\. Zhao, Y\. Jiao, N\. B\. Carlisle, H\. Xie, G\. A\. Fonzo, and Y\. ZhangMulti\-modal cross\-domain self\-supervised pre\-training for fMRI and EEG fusion\.Neural Networks184,pp\. 107066\.External Links:[Document](https://dx.doi.org/10.1016/j.neunet.2024.107066)
- Wuet al\.\(2026\)Y\. Wu, Z\. Deng, W\. Chen, J\. Li, Y\. Jiang, J\. Chen, Z\. Huang, Q\. Liu, J\. Tang, J\. Wei,et al\.LifeSide: benchmarking agents as lifelong digital companions\.arXiv preprint arXiv:2606\.04660\.
- Wuet al\.\(2025\)Y\. Wu, Y\. Peng, J\. Yu, X\. Liu, Z\. Yan, K\. Lin, W\. Su, B\. Qu, R\. Lee, and D\. YangBeyond regularity: modeling chaotic mobility patterns for next location prediction\.arXiv preprint arXiv:2509\.11713\.
- Xavieret al\.\(2025\)M\. Xavier, I\. Esteves, J\. Jorge, R\. Abreu, A\. Giraud, S\. Sadaghiani, J\. Wirsich, and P\. FigueiredoConsistency of resting\-state correlations between fmri networks and eeg band power\.Imaging Neuroscience3,pp\. IMAG–a\.
- Yanget al\.\(2023\)C\. Yang, M\. B\. Westover, and J\. SunBIOT: cross\-data biosignal learning in the wild\.arXiv preprint arXiv:2305\.10351\.Cited by:[Experimental Setup](https://arxiv.org/html/2608.29304#Sx3.SSx1.p3.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.19.1),[Table 1](https://arxiv.org/html/2608.29304#Sx3.T1.1.1.3.2)\.
- Yaoet al\.\(2025\)W\. Yao, Z\. Lyu, M\. Mahmud, N\. Zhong, B\. Lei, and S\. WangCATD: unified representation learning for EEG\-to\-fMRI cross\-modal generation\.IEEE Transactions on Medical Imaging\.External Links:[Document](https://dx.doi.org/10.1109/TMI.2025.3550206)Cited by:[Introduction](https://arxiv.org/html/2608.29304#Sx1.p2.1)\.

\*