PaperlayerView exact revision on arXiv

Source-linked AI summary

Separating Diagnosis from Disease Representation: Dual-View EEG Learning with Neural-Dynamics-Guided Deformation

Jiaying Wang, Shouqian Shi, Yutong Chen, Xu Yang, Jie Chen, Xingyu Pan, Lei Zhang, Sheng Zhong

arXiv:2609.32483v1cs.AI

TL;DR

EEG neuromodulation needs a subject-specific state that locates deviations by brain region, frequency, and lag, while existing approaches separate diagnostic evidence from anatomical readability. DMD-EEG keeps a fixed scalp expert and learns a source-space deformation of a healthy-dynamics prior, combining the experts only at the decision level. Fusion matches the strongest single expert on MDD and FEP and exceeds the source branch on PD; the source state is a coordinate-readable predictive representation with reported cross-montage rank transfer for MDD.

  • Problem

    EEG methods separate strong sensor-space diagnosis from anatomical source representations, leaving no stable coordinate system for subject-specific neural deviation in the prediction itself.

    Source: §1 · pp. 1–2

  • Method

    DMD-EEG combines a fixed scalp diagnostic expert with a source expert that learns a low-rank sparse iterative deformation of a healthy-dynamics state in shared ROI, frequency, and lag coordinates.

    Source: §1 · p. 2, §3.1 · p. 4, §3.3 · p. 4

  • Results

    Decision-level fusion matches the best branch within seed noise on MDD and FEP and exceeds the source branch on PD, while the MDD source state transfers by rank to a different-montage cohort.

    Source: §4.2 · p. 6, §6 · p. 9

  • Takeaways & Limitations

    The source state is an anatomically coordinated predictive representation whose highest-saliency coordinates align with reported disease circuitry and can support hypothesis generation.

    Source: §1 · p. 2, §6 · p. 9

    Small cohorts and sLORETA constrain localization to mesoscale cortical coordinates, excluding basal-ganglia subnuclei and not establishing synaptic delays or a measured causal pathway.

    Source: §5.2 · p. 9

Abstract

from arXiv · show

Electroencephalography (EEG)-based closed-loop neuromodulation calls for a subject-specific structured state, as opposed to a single disease probability, specifying which brain regions are deviant, at which frequencies, and at which lags. Sensor-space models keep the strongest diagnostic evidence without anatomy, source-space models give anatomy at a loss of predictive signal, and post-hoc attributions stay outside the prediction. We separate the two instead of forcing them into one representation, and propose DMD-EEG (Dual-view Multiscale Deformation for EEG), which keeps a fixed scalp spectral expert for diagnosis and models the source-space disease-related representation as a low-rank, sparse, iterative deformation of a healthy neural-dynamics prior in a $46$-region-of-interest (ROI) $\times$ $5$-frequency $\times$ $4$-lag (autocorrelation-timescale) space. The two experts meet only at a fixed decision level, so the source state is architecturally separate from the scalp expert. Across major depressive disorder (MDD), first-episode psychosis (FEP), and Parkinson's disease (PD), decision-level fusion matches the strongest single expert on MDD and FEP and exceeds the source branch on PD. On FEP the source expert is the strongest branch, the task where the deformation contributes most. The source state is an explicit ROI-frequency-lag attribution defined in a shared source coordinate system across montages, which we treat as an anatomically-coordinated predictive representation whose coordinates are directly readable and hypothesis-generating. The highest-saliency coordinates align with established disease circuitry (fronto-limbic-temporal regions in MDD, motor-cortex beta in PD), and the MDD state transfers by rank to an unseen cohort recorded with a different montage.

1 Introduction

The paper separates diagnostic fidelity from anatomical readability, addressing the lack of a method that preserves strong sensor diagnosis while making an independent source-space disease representation part of the prediction. It proposes two experts combined at the decision level and evaluates the approach across three clinical tasks.

1 Introduction

  • The central gap is a stable coordinate system for subject-specific neural deviation that preserves diagnostic evidence while locating deviation by region, frequency, and temporal relationship.Existing feature-level fusion leaves sensor and source contributions unidentified, while post-hoc attributions do not make disease-related signal an explicit coordinate of prediction.Source: §1 · pp. 1–2
  • Two separate experts combine at a low-capacity decision level: a fixed scalp spectral classifier and a source expert that iteratively deforms a subject-conditioned healthy-anchored state.The branches share no intermediate learned features or training gradients, and the source state uses explicit ROI, frequency, and lag coordinates.Source: §1 · p. 2
  • Fusion matches the best single expert within seed noise on MDD and FEP and exceeds the source branch on PD by +0.176 paired ∆AUC.The FEP source expert is the strongest branch; the fusion halves its across-seed standard deviation, while the source state's highest-saliency entries align with MDD and PD circuitry.Source: §1 · p. 2
  • The source state is presented as an anatomically coordinated, hypothesis-generating predictive representation, with MDD rank-transferable across acquisition systems.Its reported localization is supported by sensitivity attribution, ROI occlusion, and concordance with disease literature.Source: §1 · p. 2

2 Related Work

Prior EEG work includes sensor- and source-space recognition, connectivity analysis, and attribution or normative approaches, but these lines leave the paper's target combination unresolved. DMD-EEG instead learns a disease-related source deformation under a healthy-dynamics prior while retaining it for downstream analysis.

2.1 Sensor- and source-space EEG for brain disorder recognition

  • Sensor-level recognition spans MDD, FEP, and PD, but recordings, feature definitions, and evaluation protocols vary substantially across studies.Feature-level sensor–source fusion can leave each representation's contribution ambiguous (Park et al., 2025; Locatello et al., 2019).Source: §2.1 · p. 2

2.2 Directed and frequency-specific brain communication

  • Directed and frequency-specific connectivity methods estimate graphs for classification or group comparison, whereas this paper learns a disease representation as a deformation state under a healthy-dynamics prior.The cited methods include Granger causality, directed transfer function (Kamiński & Błinowska, 1991), partial directed coherence, and phase transfer entropy.Source: §2.2 · pp. 2–3, Figure 1 · p. 3

2.3 Counterfactual, interpretable, and neurobiologically validated modeling

  • Post-hoc attribution leaves a forward disease representation undefined, while the proposed state is directly supervised by disease labels and defined on explicit neurophysiological coordinates.The paper distinguishes this discriminative state from an estimated healthy-counterfactual deviation and checks state stability, coordinate readability, and concordance with known circuitry.Source: §2.3 · p. 3

3 Method

The method keeps a fixed scalp diagnostic expert separate from a source expert that learns a sparse deformation of a healthy-dynamics state. The source state retains shared ROI, frequency, and lag coordinates, and both experts meet only at fixed decision-level fusion.

3.1 Overview

  • A fixed scalp expert and a source expert produce separate probabilities that are combined by fixed averaging, keeping the scalp branch purely diagnostic.A learnable fusion weight is evaluated only as a secondary ablation.Source: §3.1 · p. 4

3.2 ROI atlas design

  • The label-independent atlas uses 46 ROIs chosen to balance transdiagnostic coverage against the resolution supported by scalp-EEG source reconstruction.Deep structures such as basal ganglia are excluded because routine scalp EEG and sLORETA cannot resolve their subnuclei.Source: §3.2 · p. 4

3.3 Healthy-dynamics pretraining

  • Healthy-only pretraining builds a source-space prior with ROI, band, and autocorrelation-lag coordinates through local self-prediction and cross-ROI communication reconstruction.The state has 46 × 5 × 4 coordinates, with autocorrelation timescales rather than inter-regional transmission delays.Source: §3.3 · p. 4

3.4 Neural-dynamics-conditioned iterative disease deformation

  • The source state is a K-step deformation of a subject-conditioned healthy anchor, constrained to shared low-rank disease directions and updated with stepwise thresholding and top-m sparsification.The top-m budget is fixed at 460 of 920 coordinates; the final deformation carries the classification signal and remains comparable across subjects and disorders.Source: §3.4 · pp. 4–5

3.5 Independent sensor diagnostic expert

  • A fixed logistic regression maps scalp power-spectral and other band-feature summaries to the sensor diagnostic probability.Features include absolute, relative, and log band power, spectral entropy, and Hjorth descriptors.Source: §3.5 · p. 5

4 Experiments

Under strict subject-level evaluation, the fixed decision-level fusion matches the best single expert on MDD and FEP and exceeds the source branch on PD. The analyses assess deformation, coordinate-level attribution, and transfer of the MDD ranking to an unseen montage.

4.1 Experimental Setup

  • The three clinical tasks use strict subject-level holdouts across three seeds, with a unified protocol for reproduced baselines and shared 46-ROI source reconstruction.The holdouts are excluded from training, early stopping, and selection; the reported values are mean ± standard deviation over three seeds.Source: §4.1 · pp. 5–6

4.2 Main Results

  • The PD fusion-minus-source AUC gain is +0.176 with a 95% CI excluding zero; scalp PSD reaches AUC 0.803, while source probes score 0.640 and 0.587.The text attributes this pattern to spectral signal and notes that sLORETA cannot resolve deep basal-ganglia pathology.Source: §4.2 · pp. 6–7

4.3 Analysis

  • The source state is independently classifiable, and its linear readout direction is cross-seed stable for MDD and FEP in two of three seed pairs.Signed cosine is 0.971 for MDD and 0.840 for FEP, with p<0.001 for both.Source: §4.2 · p. 7

4.2 Main Results

  • Fusion matches the best single branch within seed noise on MDD and FEP, and exceeds the source branch on PD (MDD 0.700, FEP 0.789, PD 0.780).Table 3 reports the dual-view results.Source: §4.2 · p. 6, Table 3 · p. 8

4.3 Analysis

  • Decision sensitivity and between-group state differences measure different properties; the highest-sensitivity ROIs qualitatively align with disease-relevant circuitry and remain hypothesis-generating.The paper reports distinct MDD, PD, and FEP circuitry patterns and cautions that the PD results do not establish elevated beta power in every source ROI.Source: §4.3 · p. 8
  • The MDD source branch transfers to a different-montage cohort with ranking AUC 0.724 ± 0.042, but a fixed 0.5 threshold misses the ordering.The transfer evidence is therefore confined to ranking.Source: §4.3 · p. 8

4.4 Ablation Study

  • FEP holdout AUC rises from 0.678 (K=0) to 0.794 (K=3), while removing healthy pretraining lowers PD AUC from 0.638 to 0.415.The no-deformation control isolates the deformation; removing multiscale temporal context also degrades MDD ranking while BAcc stays nearly unchanged.Source: §4.4 · p. 9

5 Discussion

DMD-EEG separates diagnostic fidelity from anatomical representation, combining the scalp and source experts only at a fixed decision level. Its anatomical claims are mesoscale and hypothesis-generating, with coordinate-level validation still needed in larger cohorts.

5.1 The value of separation

  • Decision-level fusion matches the best single branch within seed noise on MDD and FEP and raises PD AUC from 0.638 to 0.780.The anatomically coordinated sensitivity profile is cross-seed reproducible, and the reported sensor–source correlation on FEP is 0.25 as the source branch strengthens.Source: §5.1 · p. 9

5.2 Limitations

  • Small cohorts and sLORETA constrain localization to 46 mesoscale ROIs, excluding basal-ganglia subnuclei; the exact coordinate map remains directional pending validation in larger cohorts.Adjacent-ROI differences do not support millimeter-level localization.Source: §5.2 · p. 9

6 Conclusion

DMD-EEG separates scalp-based diagnosis from an anatomically coordinated source-space disease representation. Fusion matches the best branch on MDD and FEP and exceeds the source branch on PD; the source attribution aligns with known circuitry and transfers by rank for MDD to a different-montage cohort.

6 Conclusion

  • Decision-level fusion matches the best branch on MDD and FEP within seed noise and exceeds the source branch on PD.The source branch is strongest on FEP; the conclusion reports its direction as reproducible on MDD and FEP.Source: §6 · p. 9
  • The source branch represents disease-related deviations in explicit ROI–frequency–lag coordinates as a healthy-dynamics-conditioned low-rank deformation.The resulting attribution is concordant with known disease circuitry and, for MDD, transfers by rank to a cohort recorded with a different montage.Source: §6 · p. 9
  • The authors present the reproducible, anatomically indexed state as a structured representation that future neuromodulation work could build on.Source: §6 · p. 9

A Evaluation protocol details

The protocol uses disjoint subject-level holdouts, inner-fold selection, and explicit controls for unstable trajectories and cross-ROI leakage. Preprocessing and model-selection choices are specified alongside the deformation solver’s role.

A.1 Leak-free multi-seed holdout

  • Each of three seeds uses a disjoint, stratified 20% subject holdout, evaluated once after subject-level fold selection; results are mean ± standard deviation across three holdout evaluations.No holdout subject or clip enters training or validation, and bootstrap significance testing is conducted at the subject level, independently of seed count.Source: §A.1 · p. 14

A.2 Model selection metric

  • Fold checkpoints use a subject-level metric, defaulting to negative log-likelihood, while inner-best checkpoints and K/r settings are selected by inner-validation subject AUC.The per-disease metric setting can be suboptimal for FEP, where clip-level selection can be preferred.Source: §A.2 · p. 14

A.3 Trajectory reporting rules

  • Only stable-accumulating or final-layer-selective trajectories are eligible for discussion; oscillatory or reversing trajectories are withheld from biological interpretation.Layer reporting includes step-direction similarity, deformation magnitude, saturation, and sign flips across iterations.Source: §A.3 · p. 14

A.4 Deformation solver as an algorithm

  • The deformation algorithm iteratively updates the source state using the source representation, healthy anchor, and current deformation; the disease label enters only the outer objective.The stated outputs are the final deformation state and diagnosis.Source: §A.4 · pp. 14–15

B.1 Signal preprocessing

  • Signals are filtered from 0.5–70 Hz with a 50 Hz notch, common-average rereferenced, and reconstructed with sLORETA into a shared 46-ROI atlas.Healthy pretraining uses seven corpora totaling 269 subjects, with up to 20 randomly sampled 10-second segments per subject; preprocessing parameters are fixed before splits.Source: §B.1 · p. 15

B.2 Leakage control in cross-ROI analyses

  • Symmetric orthogonalization removes zero-lag common-driver components before cross-ROI quantities are recomputed, and both raw and corrected values are reported.Raw cross-ROI predictive increments disappear after correction, so increments alone are not treated as evidence of directed communication.Source: §B.2 · p. 15

C Baseline reproduction details

Learning-based baselines follow the proposed model’s subject-level holdout and evaluation protocol, with implementation choices fixed per baseline. Several MDD and FEP baselines show threshold degeneracy at these cohort sizes despite AUC above 0.5.

C Baseline reproduction details

  • All learning-based baselines use the same non-overlapping multi-seed subject holdout and subject-level aggregation, with holdout subjects excluded from training and model selection.Source: C Baseline reproduction details · p. 15
  • EEGNet is trained from random initialization on native-montage four-band signals, while CBraMod uses its official checkpoint and fine-tuning pipeline.DeeperBrain follows its publication protocol because its training code was unavailable; reproduction choices are fixed per baseline and recorded with shared settings.Source: C Baseline reproduction details · p. 15
  • On MDD and FEP, several learning-based baselines predict a single class, yielding balanced accuracy of 0.5 with zero variance, while their AUC remains above 0.5.The paper characterizes this as threshold degeneracy at these cohort sizes, distinct from numerical failure.Source: C Baseline reproduction details · p. 15

D Scalp versus source spectral discrimination

The section compares scalp and source representations with matched PSD and EEGNet probes, using source-minus-scalp AUC as the penalty. It also specifies the 46-ROI atlas boundary and reports the model’s parameter count.

D Scalp versus source spectral discrimination

  • The PSD probe and EEGNet are each applied to scalp and source-ROI representations, and their within-probe source-minus-scalp AUC penalty tests whether higher capacity closes the gap.A negative penalty means the source representation is weaker; the table reports subject-level mean AUC.Source: D Scalp versus source spectral… · pp. 15–16, Table 4 · p. 16

E.1 ROI atlas (46)

  • The atlas contains 46 separate left/right ROI coordinates, while basal-ganglia subnuclei are excluded as primary ROIs because routine scalp EEG with template-head sLORETA does not support independent subnuclear claims.Source: §E.1 · p. 16

E.2 Hyperparameters

  • The source branch has 168k trainable parameters, and the total DMD-EEG model has approximately 169k parameters.Disease-specific totals are MDD 169,628, FEP 168,623, and PD 169,592; the gnorm development readout is excluded.Source: §E.2 · p. 16

F Ablations

The ablations examine deformation depth and rank, healthy pretraining, fusion weighting, and paired fusion gains. They support a healthy-dynamics prior and fixed averaging, while the PD fusion gain has a confidence interval excluding zero.

F Ablations

  • The selected deformation settings were MDD K=4, r=1; FEP K=3, r=1; and PD K=2, r=2, with controls below selected depth on every task.The small scan and shared ROI–band–lag space were chosen for cohorts differing in size, montage, and sampling rate; the inner-validation–holdout gap reflects selection among configurations on a small inner pool.Source: F Ablations · p. 16
  • Clearing the healthy-pretrained checkpoint reduced holdout AUC on all tasks, most sharply for PD, from 0.638 to 0.415.Holdout AUC also fell from 0.794 to 0.778 on FEP and from 0.622 to 0.600 on MDD.Source: F Ablations · p. 16

F.1 Fusion variants

  • Neither adaptive fusion alternative reliably improved over the fixed average, and selected α often collapsed to 0 on the small inner pool.The fixed α=0.5 average was retained, preserving the source state’s structural independence.Source: §F.1 · p. 17

G Paired bootstrap of the fusion gain

  • The PD fusion-minus-source gain had a 95% confidence interval that excluded zero in the paired subject-level bootstrap.The bootstrap used 20,000 resamples of pooled holdout subjects with the same resampling index for both branches.Source: G Paired bootstrap of the fusion gain · p. 17

H Fixed five-fold reference

The fixed five-fold reference reports source-side readout results, while additional diagnostics test pathway contribution, source-model components, and cross-ROI predictive increments. These analyses distinguish predictive communication from biological directionality and are secondary to the strict-holdout generalization estimate.

H Fixed five-fold reference

  • The best source-side configuration was the per-subject logit average of direct-deformation and communication-pathway readouts, with MDD 0.763/0.712 and FEP 0.683/0.693.These fixed subject-level five-fold values are selection-biased; the multi-seed holdout protocol is the stated generalization estimate.Source: H Fixed five-fold reference · p. 17

I.1 FL-DPD communication-pathway ablation

  • Removing pathway-fusion weights reduced subject-level accuracy from 0.611 ± 0.019 to 0.491 ± 0.046 and AUC from 0.666 ± 0.016 to 0.500.This development comparison supports a predictive role for the pathway, but leaves direction and lag semantics untested.Source: §I.1 · p. 18

I.2 Component-level diagnostics of the source model

  • In early healthy-pretrained variants, no single edge was significant under subject-level permutation (0/462), while observed AUC was 0.69–0.74 versus null 0.458±0.055.The sparse gate stayed at its initialization, the cross-frequency term near zero, and the source-side signal was diffuse across edges and strongest in spectral power.Source: §I.2 · p. 18

I.3 Leakage correction for cross-ROI predictive increments

  • Cross-ROI incremental predictive ability of +0.0116 to +0.0144 vanished after symmetric orthogonalization (−0.0003), supporting a predictive—not biological-directionality—interpretation.The top-1 leakage component explained 46–52% of variance and the top-3 explained about 100% of the increment.Source: §I.3 · p. 18

I.4 Cross-task state dependence of MDD deviations

The cross-task MDD state comparison did not establish a verified resting-versus-ERP dissociation; transfer and supplementary analyses further delimit where source-state results hold. The supplementary figures distinguish descriptive state maps from unbiased performance estimates.

I.4 Cross-task state dependence of MDD deviations

  • 6/462 resting and 7/462 ERP frequency edges were nominal; neither yielded FDR-significant edges, and the pathway-by-state interaction was nonsignificant (t=−0.985, p=0.329, d=−0.269, n=53).Masking predefined DMN or fronto-limbic edge sets changed AUC by less than 0.01, supporting preliminary state dependence rather than a verified dissociation.Source: §I.4 · pp. 18–19

I.5 External cross-cohort transfer

  • 0.724±0.042 AUC for MDD transfer contrasts with 0.542±0.032 for FEP; MDD ranking transfers, while FEP discrimination is cohort-specific.The fixed 0.5 threshold mislabels many MDD subjects (BAcc 0.548 ± 0.073); only ranking AUC is used as transfer evidence because cohorts are small and folded metrics are class-balance sensitive.Source: §I.5 · pp. 19–20

I.6 Additional ablations

  • Temporal position embeddings and attention pooling showed no reliable contribution; attention versus mean pooling changed AUC by +0.013 ± 0.027.In the FEP fine-tuning grid, degradation traced to focal-loss exponent change (γ=2 →0), with fold-1 AUC changing from 0.833 to 0.500; communication-loss weight stayed unchanged.Source: §I.6 · p. 20

I.7 Lag-axis validity: synthetic recovery and lag perturbation

  • 1.00 accuracy on both latent and raw ACFs recovered planted lags, against chance 0.25.This positive control tests whether the four-lag representation can encode a known lag.Source: §I.7 · p. 21, Table 12 · p. 21
  • −0.160, −0.138, and −0.086 ∆AUC followed replacing per-τ ACF with time means for MDD, FEP, and PD; lag added beyond band for MDD and PD, but not FEP (+0.006).Permuting or rolling τ anchoring produced smaller drops (−0.03 to −0.06); the ACF is even in τ, so the lag axis encodes |τ| rather than direction.Source: §I.7 · p. 21, Table 13 · p. 21

I.8 Attribution coordinates / I.9 ROI occlusion on the decision

  • 0.83–0.87 cross-seed axis cosine marks stable ROI saliency, concordant with occlusion rankings (ROI Spearman +0.31 MDD, +0.59 PD).Table 15 ranks single ROIs by occlusion magnitude; joint-set results are exploratory, and the nominal permutation p-values are uncorrected for searching 46 ROIs, making the table descriptive.Source: §I.9 · pp. 21–22, Table 14 · p. 22, Table 15 · p. 23

J.1 ROI × frequency × lag effect maps / J.2 Per-subject deformation trajectory and state space

  • −0.02 to −0.11 cell-level correlation separates patient–HC state-difference maps from decision-sensitivity maps; trajectories depict deviation accumulating through iterative solving.The maps use directly readable ROI, band, and real per-τ ACF lag axes; only AUROC panels provide unbiased performance estimates.Source: §J.1 · p. 22, §J.2 · p. 22, J Supplementary figures · p. 22

J.3 Dual-view AUROC / J.4 Confidence scatter / J.5 Disease pathway glass brains

  • Fusion matches the strong source branch on FEP and exceeds it on PD while matching the sensor; glass-brain sensitivity and patient–HC offset encode distinct quantities.Confidence scatterplots show branch-specific disagreements, while glass-brain pathway arrows are illustrative rather than learned connectivity.Source: §J.4 · p. 22, §J.5 · p. 23, Figure 5 · p. 26
Loading 2609.32483v1…