acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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 -region-of-interest (ROI) -frequency -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.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, interpretable representation learning, multi-view learning, source-space EEG, neural dynamics, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_X9ptOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DMD-EEG, with independent branches for scalp spectral diagnosis and source-space representation learning, combined through decision-level fusion. The source branch uses healthy-dynamics pretraining and low-rank sparse iterative deformation and is evaluated on MDD, FEP, and PD.

Critical strengths

S1. The independent sensor/source branches preserve scalp-level diagnostic information while providing source-space representations with explicit anatomical and spectral coordinates.

S2. The evaluation covers three diseases, uses strict subject-level holdout, and includes module ablations, attribution stability checks, leakage controls, and external transfer analyses.

Critical weaknesses

W1. The link between the biological rationale and the computational assumptions underlying the healthy-dynamics prior and low-rank iterative deformation needs further analysis.

W2. The methodological contribution beyond existing methods needs clarification.

Questions for authors

Q1. Pretraining on healthy data and preserving spectral and temporal structure are well motivated. Could the authors clarify which properties of the healthy-dynamics prior are grounded in neurophysiological research while which are modeling choices? What features of neural activity are local prediction, cross-ROI prediction, and autocorrelation each intended to preserve?

Q2. Could the authors add a conceptual comparison with existing sensor–source fusion, EEG normative modeling, and sparse iterative representation methods?

Q3. Could the authors explain how the low-rank parameterization, with a shared disease direction and subject-specific amplitudes, balances regularization in small samples with individual variability?

2 comments
AuthorsOpenReview ↗

Response to X9pt: physiological grounding, related-work positioning, and low-rank parameterization

X9pt

We thank the reviewer for the three constructive questions on the physiological basis of the prior, its relation to existing methods, and the low-rank parameterization. We answer each below, with revisions(.pdf) for your points marked in green.

W1/Q1 : physiological grounding versus modeling choices in the healthy prior

Answer. Oscillatory activity and its temporal correlations are estimable EEG properties, while frequency-specific inter-regional interactions are well documented. The 46-ROI atlas, five-band partition, and four lags (τ=1,2,4,8\tau=1,2,4,8) are modeling choices fixed before training, while each subject supplies the coordinate values. Each objective targets a distinct structure: local self-prediction preserves within-ROI short-time and cross-band dynamics, cross-ROI prediction preserves inter-regional predictive structure, and the ACF axis preserves each ROI–band correlation-decay timescale. Disease labels then fine-tune the encoder and train the deformation and readout.

Where changed. §3.3 "Oscillatory activity and its temporal correlations are estimable EEG properties..."; §2.2 (healthy-prior definition).

W2/Q2 : relation to sensor–source fusion, normative modeling, and sparse iterative representations

Answer. Unrolled sparse coding (LISTA) and EEG graph unrolling optimize a signal-processing quantity, not a disease state on anatomical coordinates (§2.1, §3.4). Mixing sensor and source features inside one feature vector leaves the source contribution ambiguous, whereas decision-level fusion keeps an inspectable ROI–frequency–lag deformation (§2.1, §3.1). Normative modeling quantifies deviation from a healthy reference, whereas we learn a disease-supervised deformation from a healthy-pretrained anchor (§2.3, §3.3). The modeled quantity and clinical target differ. Prior work models a connectivity deviation for treatment-response prediction, while ours models a structured deformation for disease classification.

Where changed. §2.1 (LISTA/graph unrolling; Shim sensor–source concatenation); §2.3 (Tong normative modeling).

Q3 : low-rank parameterization(regularization versus individual variability)

Answer. The parameterization separates what is shared from what is individual. Each disease learns one cohort-shared basis VD∈Rr×920V_D\in\mathbb{R}^{r\times920}, which defines rr disease axes and fixes the directions all subjects move along. This provides the regularization. The number of free axes does not grow with the cohort, and only rr directions are estimated from a small sample. Individual variability enters in the coordinates, not the axes. For each subject, a small amplitude network predicts per-step signed coefficients sk(x)∈Rrs_k(x)\in\mathbb{R}^{r} from the subject-conditioned anchor h0(x)h^0(x), and the update is Uk(x)=VD⊤sk(x)U_k(x)=V_D^\top s_k(x), so subjects differ in how far and in which direction they move along the shared axes. The top-mm support (m=460m{=}460 of 920920 coordinates) is ranked from the subject's current deformation and re-selected at every step, so the retained subset adapts per subject. We also test the capacity of the shared low-rank basis by varying only the rank rr while holding solver depth and sparsity fixed for each disease (Table 10). Across all three tasks, holdout AUC is highest at a small rank (r≤2r\le2) and does not improve at larger ranks. On FEP the r=1r{=}1 basis is strongest at AUC 0.7940.794, and larger ranks do not recover the gap: r=2r{=}2 reaches 0.7170.717, r=4r{=}4 falls to 0.6940.694 (paired Δ\DeltaAUC −0.137-0.137, 95% CI [−0.277,−0.011][-0.277,-0.011]), r=8r{=}8 reaches 0.7110.711, and a per-coordinate control without the shared basis reaches 0.7220.722. The larger variants therefore add parameters without a resolved gain. On MDD, r=1r{=}1 and r=2r{=}2 tie at 0.6220.622, so the higher rank brings no benefit. On PD, r=2r{=}2 gives the best AUC (0.6380.638), while r=1r{=}1 is unstable (0.460±0.2430.460\pm0.243) and r=4r{=}4 drops to 0.3930.393. Readability of the resulting state is not inferred from this sweep and is assessed separately through cross-seed stability and occlusion (§4.3).

Where changed. §3.4 "VDV_D limits shared update capacity, while sk(x)s_k(x) and top-mm retain individual variation..."; §4.4 "We also test the capacity of the shared low-rank basis..."; Table 10.

Thanks for the response and revisions, which have addressed my concerns regarding the healthy-dynamics prior, methodological positioning, and low-rank parameterization. I will raise my rating accordingly.

Reviewer_5RxNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DMD-EEG to combine a sensor-space models for diagnostic evidence and a source-space models for anatomically interpretable representation from the resting-state EEG. The sensor-space expert classifies scalp spectral features, while the source-space expert learns supervised low-rank, sparse deformations of a healthy pretrained representation in ROI x Frequency x lag state. Each expert branches are trained independently and then are averaged in final decision. The method is evaluated on MDD, FEP, PD.

Critical strengths
  1. Clear problem statement and distinction between diagnostic accuracy and anatomical interpretability is insightful.

  2. Proposed method matches with the problem statement. The paper reports each expert branch and the fusion in steps, so the performance of each experts and the effect of the fusion can be evaluated.

  3. The evaluation is not only done in classification accuracy, but extends to sensitivity analysis, ROI occlusion, and lag perturbation, which examine what affects predictions. and also source model is tested on other cohorts.

Critical weaknesses
  1. The motivation of separating the two expert branches is well-stated. but the methodological contribution of just simply averaging two path in final decision itself seems to be limited.

  2. The benefit of fusioning, averaging two individual experts varies across tasks and seems to be weak. Mean AUC increases on MDD, but the fusion result does not show explicit gain over the stronger individual expert, on FEP and PD. The improvement is also agains the source branch on PD. So evidence for preserving the diagnostic performance while preserving the source path is still unclear.

  3. The contribution of low-rank constraint is not sufficiently evaluated. Table 7 compares ranks 1 and 2 at fixed depth on MDD, but the mean AUC stays the same. There is no high rank or unconstrained comparison, and the settings change in PD dataset. So the current ablations do not clearly show how the low-rank constraint helps generalization and interpretability.

Questions for authors
  1. Other than independently training and averaging 2 expert paths, what is the main contribution of the dual-view design? Since the benefit of the fusion differs across the tasks, in what cases does it help over the stronger expert?

  2. Is there any systematic ablations to validate the contribution of the low-rank constraint? and is there comparisons to higher-rank or unconstrained models while fixed depth or sparsity?

  3. Is there any computational comparison(Flops, memory, inference latency, etc) of the full method, including the source reconstruction?

  4. What were the criteria for selecting the 3 tasks and the baselines? Since simple spectral baselines also show competitive results, why are these tasks suitable for evaluating the benefits of the proposed method?

3 comments
AuthorsOpenReview ↗

Response to 5RxN: dual-view separation, fusion vs. stronger branch, rank sweep, cost, and task/baseline rationale

5RxN

We thank the reviewer for recognizing the separation of diagnostic and anatomical outputs and the branch, occlusion, and cross-cohort analyses. We address the five points in turn, with revisions(.pdf) for your points marked in blue.

W1/Q1 : contribution of the dual view beyond averaging two experts

Answer. The contribution is the separation, not the average. A fixed low-capacity sensor path preserves diagnostic evidence, while an independently trained source path learns a disease state in explicit ROI×band×lag coordinates. The branches share no intermediate features or gradients, so the source state is not absorbed by fusion and remains independently classifiable and inspectable. Fixed averaging is only the decision interface that keeps both outputs. Evidence: the two predictions are weakly to moderately correlated and cover different errors (Table 3; Figure 6), the source branch is strongest on FEP, and on PD the sensor carries diagnosis while the source remains complementary.

Where changed. §1 contribution "Dual-view separation..." and "We average the independently trained probabilities..."; §5.1 "Diagnosis and disease representation have different requirements..."; §4.2 "The two views cover different errors...".

W2/Q1 : when the fusion helps over the stronger expert

Answer. The fusion matches the stronger branch on MDD and FEP: the paired ΔAUC against it is +0.052 on MDD and +0.030 on FEP, and both intervals span zero. On FEP the source branch is the strongest single expert (0.794), and the fusion stays close (0.789) while halving its across-seed SD (0.139→0.055). On PD the fusion exceeds the source branch with a CI excluding zero (+0.176, [+0.048,+0.308]) while matching the sensor; it also has the highest F1_pos (0.792 vs 0.784 sensor and 0.783 source) and AUC (0.780 vs 0.772). The source branch recovers 11 patients the sensor misses, against 6 that the sensor alone recovers (Figure 6), and the fusion raises MDD AUC from 0.622 to 0.700. The contribution is therefore the separation: an independently classifiable, anatomically indexed source state that fusion does not absorb, with averaging as the low-capacity interface that keeps both outputs.

Where changed. §4.2 "Table 3 reports the dual-view results..." and "matches the stronger branch..."; §5.1 "matches the stronger branch"; §5.2 limitation; §6 conclusion; Table 12; Figure 10; Figure 6.

W3/Q2 : low-rank constraint across higher ranks and an unconstrained control

Answer. We also test the capacity of the shared low-rank basis by varying only r at fixed depth and sparsity for each disease (Table 10). Holdout AUC is highest at a small rank (r≤2) on all three tasks and does not improve at larger ranks. On FEP, r=1 is strongest (AUC 0.794), r=4 falls to 0.694 (paired ΔAUC −0.137, 95% CI [−0.277,−0.011]), and the larger r=8 and per-coordinate controls add parameters without a resolved gain. On MDD, r=1 and r=2 tie at 0.622. On PD, r=2 is best (0.638), while r=1 is unstable (0.460±0.243) and r=4 drops to 0.393. Readability is assessed separately through cross-seed stability and occlusion (§4.3).

Where changed. §4.4 "We also test the capacity of the shared low-rank basis..."; §3.4 "VDV_D limits shared update capacity..."; Appendix Table 10; §4.3 for stability and occlusion.

Q3 : computational cost including source reconstruction

Answer. Offline sLORETA reconstruction dominates end-to-end runtime, while the online forward takes about 2.1–2.3 s per subject. At batch 8 the online source forward peaks at 9148 MB of GPU memory, and a batch-1 source-only clip uses 1152 MB. One representative FEP forward (seed-42, K=3) costs 227.14 GFLOPs source-only and 255.75 GFLOPs with FL-DPD. The encoder is shared across tasks, while solver depth differs slightly. The sensor LR and fusion add 0.38–0.39 ms and 0.005–0.006 ms per decision.

Where changed. Appendix "Full-pipeline cost including source reconstruction"; Table 11; Figure 12.

Q4 : task and baseline selection

Answer. Tasks were chosen because they provide public subject-level labels and 60–128-channel EEG(fewer channels yield a larger source-reconstruction error) across three phenotypes (Table 1), namely MODMA (mood, 24/29), ds003944 (psychotic, 50/32), ds004584 (movement, 100/49). All three reconstruct to the same 46-ROI atlas under one subject-level holdout protocol, a cross-disease comparison the datasets previously lacked. Baselines span strong simple spectral, general EEG, and task-specific methods. Simple spectral baselines being competitive fits the division of labor. Scalp spectra answer whether a subject is a patient, while the source state addresses where the deviation lies.

Where changed. §4.1 Datasets and Baselines; Appendix A; baseline reproduction appendix.

Thank you for the additional experiments and clarifications. It addresses my concern about the low-rank constraint. However, my main concern was what additional value does the dual-view design adds beyond combining two independently trained models. The rationale of separation is clear, but under limited gains over stronger path, clarifying the purpose of separation alone does not address this concern. I therefore maintain my current rating.

We thank the reviewer for the careful follow-up and for recognizing that the low-rank concern has been addressed. You ask what the dual-view design adds beyond independently training and averaging two models. The methodological innovation lies in the disease-supervised source deformation and its shared, anatomically readable ROI × frequency × lag coordinates. The dual-view design keeps this classification-relevant state separate from the scalp expert and makes both available in a joint decision (Figure 1). The fixed average is a deliberate decision rule, not the proposed algorithmic novelty.

The source state offers more than a diagnostic probability. Section 4.3 reports decision sensitivity and group effects as distinct readouts on its coordinates (Figures 2–3). The ROI-level sensitivity profile is stable across seeds. Its MDD fronto-limbic-temporal and PD motor/beta patterns are qualitatively consistent with the cited literature, and masking selected ROI blocks lowers the source expert’s AUC (Tables 17–18). The source expert also makes its own predictions: on FEP, it has the highest mean AUC among the branches (0.794), compared with 0.533 for source PSD and 0.628 for source EEGNet (Tables 3–4). These results connect the anatomically readable state to the diagnostic decision.

Retaining the scalp view addresses a different requirement. Under the same PSD probe, moving from scalp to source lowers mean AUC by 0.078/0.167/0.163 on MDD/FEP/PD (Table 4), consistent with spectral evidence being diluted by source reconstruction and ROI aggregation. Figure 6 reports the numbers of held-out subjects correctly classified by the scalp expert but misclassified by the source expert, and those correctly classified by the source expert but misclassified by the scalp expert. We therefore retain both complementary views and combine only the experts’ final probabilities using fixed equal weights, leaving each readout available for separate evaluation.

The joint prediction’s observed mean AUC exceeds both branches on MDD and stays close to the stronger branch on FEP and PD (Table 3), with paired uncertainty reported in Table 12. On PD, its mean F1_pos is also slightly higher than either branch (0.792 versus 0.784/0.783). The paired AUC comparisons assess the diagnostic performance of providing both views. The source state’s contribution is supported by its independent predictions and coordinate-level analyses, not by the size of an AUC gain from averaging.

The dual-view design retains scalp diagnostic evidence that source reconstruction can weaken, while confining the scalp expert’s influence to fixed, equal-weight fusion of the final probabilities and leaving the source disease state separate and anatomically readable.

Reviewer_hC7aOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces DMD-EEG, a dual-view framework for EEG-based diagnosis. It combines a frozen logistic regression on scalp PSD features with a source-space expert that represents sLORETA activity using a 46×5×446 \times 5 \times 4 ROI, frequency, and autocorrelation-lag anchor, pretrained on healthy subjects to capture normal neural dynamics. A sparse, low-rank solver then learns subject-specific deformations along shared disease directions. These deformations feed a linear classifier, whose predictions are fused with the sensor expert's output. The learned deformations also provide interpretable attributions across MDD, first-episode psychosis, and Parkinson's disease.

Critical strengths
  1. The goal and motivation of the paper is clear and paper is very-well written and easy to follow.

  2. I like the shared, interpretable coordinate system across disorders. The paper keeps three readouts distinct: the deformation, the decision sensitivity and group effect size, this will be useful for clinical neuroscience beyond the classification accuracy.

  3. Many EEG deep learning papers omit simple baselines but authors included simple logistic regression baselines on Scalp PSD both as a baseline and as a frozen fusion.

Critical weaknesses
  1. The per-fold standard deviations (±0.10−0.14)(\pm0.10-0.14) are comparable to most of the performance gaps between methods, and the cohorts are relatively small. Could you report paired significance tests comparing DMD-EEG (Full) against Scalp PSD + LR for each task, along with subject counts and class prevalence? On PD, the Full model does not outperform LR in BAcc, while DMD-EEG-Source has the lowest AUC among all methods. What evidence supports the claim that the source expert improves diagnostic performance beyond providing interpretability?

  2. Each amplitude sk(x)s_k(x) is predicted from h0h^0 and does not depend on the current Δhk\Delta h^k. This means the KK updates essentially reduce to a single vector VD⊤c(x)V_D^\top c(x), aside from thresholding. Could you include an ablation with K=1K = 1 and without soft thresholding or top-mm selection (MLP(h0h^0) →\to rr →\to linear head)? It would also be helpful to report typical ∣Δh∣|\Delta h| values relative to τ≤10−3\tau \leq 10^{-3} to clarify whether soft thresholding has any practical effect.

  3. Since the backbone is fine-tuned using disease labels, and h0h^0 is used to predict the deformation, it is unclear how much the method actually relies on the healthy pretraining prior. Could you compare a frozen pretrained backbone, a fine-tuned backbone, and random initialization without healthy pretraining? A linear probe on h0h^0 alone would also help. Finally, can you confirm that the seven healthy pretraining datasets have no subjects or recording sites in common with the controls used for fine-tuning and testing?

  4. Several baselines appear to collapse to predicting every subject as positive (BAcc 0.500 ±\pm 0.000, F1 0.769 ±\pm 0.000 on FEP), while others perform consistently below chance (e.g., Scalp PSD + SVM AUC 0.344 ±\pm 0.042 on MDD and CBraMod AUC 0.406 on FEP). Were all baselines given comparable tuning budgets, class balancing, and data splits? Could you also clarify what might explain these below-chance results?

I believe the paper's motivation and novelty are clear, but the empirical claims would benefit from stronger ablation studies and numerical validation. I would be open to raising my score if these concerns are adequately addressed.

Questions for authors

Please see the Weaknesses section.

3 comments
AuthorsOpenReview ↗

Response to hC7a: paired fusion evidence, sparsity controls, pretraining controls, and baseline fairness

hC7a

We thank the reviewer for recognizing the shared anatomical coordinates, the three separated readouts, and the Scalp PSD+LR baseline. We address the four points in turn, with revisions(.pdf) for your points marked in red.

W1 : paired evidence for Full vs PSD+LR, and the source expert's diagnostic value

Answer. Against the sensor, the fusion is positive but not significant on MDD (+0.052, CI [−0.070,+0.173]) and FEP (+0.065, CI [−0.030,+0.164]), and essentially zero on PD (−0.006, CI [−0.049,+0.035]). Table 2 gives the PD BAcc (Full 0.683 vs LR 0.700). The cohorts are MDD 24/29, FEP 50/32, and PD 100/49, with per-seed holdouts 5/6, 10/6, and 20/10 (33/48/90 pooled). The source branch contributes diagnostic information in a task-dependent way: it is the strongest branch on FEP (0.794 vs sensor 0.700), and on PD, although weak (0.638), it recovers 11 patients the sensor misses against 6 the sensor alone recovers (Figure 6). Subject counts and per-holdout prevalence are in Appendix A.

Where changed. §4.2 "With the same PSD classifier applied to both representations..." and "Relative to the Scalp PSD+LR (sensor) branch...", which ends by pointing to "Cohort sizes and holdout prevalence are in Appendix A."; Table 2; Table 12; Figure 10.

W2 : whether K steps collapse to one update, and whether soft-threshold matters

Answer. With both operators off, the K steps collapse to one shared low-rank update because every amplitude reads the same h0h^0. With the operators on, top-m re-selects the support at each step and τk\tau_k shrinks it, making the composition nonlinear. We therefore ran a controlled comparison to test the difference directly. The rank-matched K=1 both-off control (prediction-structure-equivalent to MLP(h0h^0) → r → linear, retaining the deformation L1) trails the full solver on FEP and PD (0.717/0.393 vs 0.794/0.638). On MDD its AUC is unchanged (+0.004) but its positive-class F1 is lower (0.381 vs 0.548). The full multi-step sparse solver is therefore retained, with a task-dependent margin. The learned τ≈10−5\tau\approx10^{-5} is far below the pre-threshold median magnitude (0.040–0.139). Soft-threshold zeroes 0–0.12% of entries while top-m fixes the support at 460/920, so top-m drives the sparsity.

Where changed. §4.4 "To test whether the selected depth matters..." and "The two sparsifiers are also uneven in practice..."; the one-shot control is relabelled "K=1, with soft-threshold and top-m enabled"; Table 9; Figures 8 and 11.

W3 : frozen/fine-tuned/scratch, linear probe, and pretraining–cohort overlap

Answer. In Table 8, Source-Full (pretrained encoder, fully fine-tuned, solver and head trained) reaches AUC 0.622/0.794/0.638 on MDD/FEP/PD, and its K=0 no-deformation control, which routes h^0 to the linear head as a post-finetuning h^0 readout, gives 0.600/0.678/0.623, so the solver adds beyond the anchor. Scratch (random initialization, same protocol) gives 0.678/0.706/0.512 and, on MDD, ranks higher than Source-Full (0.678 vs 0.622) while its positive-class F1 is much lower (0.354 vs 0.548), so pretraining improves the decision. Frozen encoder (encoder fixed, scale_axis/solver/head trained) gives 0.467/0.628/0.648, below Source-Full on MDD/FEP and matching it on PD (0.648 vs 0.638), so fine-tuning the encoder matters on MDD/FEP but little on PD. Frozen freqz (encoder fixed, no solver, linear head on the encoder's frequency state, bypassing scale_axis) reaches MDD/PD AUC 0.700/0.633 with BAcc 0.500 and FEP AUC 0.428, so the frozen healthy representation carries ranking but not fixed-threshold decisions, unlike the K=0 readout, which uses the fine-tuned h^0. For overlap, given the anonymity of the public datasets, the manuscript excludes overlap at the dataset and documented-site level, not at the individual-subject level.

Where changed. §4.4 "Healthy pretraining improves the decision, since Source-Full..."; Appendix Table 8; Figure 9; §4.1 "The seven pretraining corpora and the three clinical cohorts are distinct datasets..."; §5.2 "Given the anonymity of the public datasets...".

W4 : baseline fairness and below-chance results

Answer. All methods use the same fixed 0.5 threshold, validation-AUC checkpoint selection with a BAcc tiebreak, and the same seven healthy corpora for any pretraining. The BAcc 0.500 rows are single-class threshold collapses, with AUC still carrying ranking information. The below-chance AUCs are ranking failures on few positive–negative pairs per holdout (30 for MDD, 60 for FEP), where a few misranked subjects move AUC far from 0.5.

Where changed. §4.1 "We reproduced the principal subject-level baselines..." and "Within this protocol, all methods use the same fixed 0.5 threshold..."; §4.2 "reports every baseline together with..."; Table 2; Appendix baseline reproduction.

Reviewer_hC7aOpenReview ↗

Reply to Authors

I thank the authors for the thorough response and for the transparency of the revision, including the paired bootstrap intervals, the new controls, and the explicit deletion of unsupported claims.

However, the new results largely confirm my original concerns rather than resolve them:

(1) Against Scalp PSD + LR, the full model’s paired AUC gains span zero on MDD and FEP and are essentially zero on PD, where BAcc is also lower. With holdouts of roughly 11 (MDD), 16 (FEP) and 30 (PD) subjects per seed, the evidence that the source expert improves diagnosis remains weak. The FEP advantage of the source branch is not accompanied by a paired test, and the PD discordance (11 vs. 6 subjects) is too small to be conclusive.

(2) The soft-threshold is effectively inactive, and top-m retains 50% of coordinates, so the “sparse” deformation is only weakly sparse. The K=1 control’s PD AUC of 0.393 suggests instability rather than a clean ablation, and no intervals are given for these differences.

(3) The pretraining controls are mixed: scratch training outperforms the pretrained model on MDD AUC (0.678 vs. 0.622), and the frozen-encoder variants fall below chance on MDD and FEP. The argument that pretraining helps relies on switching to positive-class F1. As a result, the contribution of the healthy neural-dynamics prior, which is central to the paper’s framing, is not convincingly established.

(4) The explanation for below-chance baselines does not account for Scalp PSD + SVM’s AUC of 0.344 ± 0.042 on MDD: the small across-seed variance indicates a systematic rather than sampling effect. It also remains unclear whether baselines received comparable hyperparameter tuning budgets.

I continue to find the motivation and the interpretable coordinate system valuable, but the empirical support for the diagnostic and pretraining claims remains insufficient. I therefore maintain my score.

We thank the reviewer for the careful follow-up and for the valuable suggestions that have helped us examine the claims and evaluation more closely. We respond to the four remaining points in turn.

1. Diagnostic value and separation. Thank you for asking for a direct paired FEP source–sensor comparison. On the held-out predictions, pooled AUC is 0.737 for source and 0.702 for sensor (ΔAUC +0.035 [−0.162, +0.235]); BAcc at the fixed 0.5 threshold differs by +0.050 [−0.116, +0.218]. The per-seed AUC differences are −0.183, +0.233, and +0.233. Table 3’s 0.794 versus 0.700 is the mean of per-seed AUCs, whereas the paired result uses pooled predictions. The source has a higher point estimate on FEP, although the direct paired difference is not statistically significant.

The branches have distinct, complementary roles: the fixed scalp expert supplies sensor-level diagnostic evidence, while the source expert supplies an independently predictive, anatomically readable ROI × frequency × lag state. We compare their diagnostic metrics to show how each performs, not because the source branch must outperform the sensor branch on every task. Their weak-to-moderate prediction correlations (Table 3) and differing error patterns (Figure 6) further illustrate this complementarity. We report the fusion comparisons and their intervals in Table 12; as §5.2 states, the intervals against the stronger branch span zero.

2. Sparsity and the K=1 control. We agree that soft-thresholding has little effect on support: it zeroes 0–0.12% of entries, while top-m retains 460/920 coordinates. Thus, top-m supplies the support selection, and the resulting deformation is moderately rather than highly sparse.

For the rank-matched, both-operators-off K=1 control cited by the reviewer, Table 9 reports paired ΔAUC intervals relative to Source-Full: +0.004 [−0.128, +0.139] on MDD, −0.035 [−0.117, +0.043] on FEP, and −0.182 [−0.354, −0.014] on PD. To examine the PD result without changing solver depth, we added a K=1, r=2 both-on control in Appendix F. Its AUC is 0.572 ± 0.107, compared with 0.393 ± 0.097 for both-off. The paired both-off-minus-both-on ΔAUC is −0.176 [−0.324, −0.022]; all three per-seed differences are negative (−0.395, −0.025, −0.115), although their magnitudes vary. This fixed-depth, fixed-rank comparison shows that retaining the operators improves the PD result of the simplified control.

3. Healthy pretraining. F1_pos was not introduced in response to the MDD AUC result: AUC, BAcc, AUPRC, and F1 are specified in §4.1 and reported together in Table 8. We agree that scratch has a higher per-seed mean AUC on MDD (0.678 versus 0.622); its paired AUC interval versus Source-Full nevertheless spans zero. At the fixed 0.5 threshold, Source-Full has higher MDD BAcc (0.622 versus 0.506) as well as F1_pos (0.548 versus 0.354). Both fixed-threshold metrics also favor Source-Full on FEP and PD; on PD, the paired scratch-minus-Source-Full BAcc difference is −0.108 [−0.205, −0.017]. We therefore report a task- and metric-dependent pretraining effect, rather than replacing the unfavorable MDD AUC with F1_pos.

The frozen results are also mixed, rather than uniformly below chance on MDD and FEP: frozen encoder reaches FEP AUC 0.628, and the frozen-frequency-state probe reaches MDD AUC 0.700, although the latter has BAcc 0.500. As §3.3 describes, the healthy-pretrained encoder supplies the initialization for subsequent disease-label fine-tuning; the frozen probes test a different condition. Table 8 shows that freezing hurts AUC on MDD and FEP, while its PD result is close to Source-Full.

4. Baseline tuning and MDD SVM. We audited the low MDD PSD+SVM AUC. On the same fitted models, margin/probability AUCs were 0.533/0.417, 0.700/0.500, and 0.567/0.333 across seeds. This implicates calibration alongside small-holdout sensitivity in the reported 0.344±0.042.

We chose SVC(probability=True) from the outset because Table 2 evaluates positive-class probabilities for AUC and applies the same outputs at a fixed 0.5 threshold for BAcc/F1. The margin is a valid SVM score, but replacing the reported probability output after inspecting holdouts would change that protocol. We retain the probability-based result regardless of which output scores higher.

For tuning, a correction to our earlier reply: class weights were fixed but method-specific (Appendix C; E.2). All methods shared subject-level splits and inner-only selection. LR/linear SVM searched C={0.1,1,10}, RF n_estimators={300,600} x max_depth={None,8}, FEP microstate SVM the same three C values, and DMD-EEG K/r. Deep baselines used fixed method-specific recipes while training or fine-tuning on each target cohort. Given the available compute, we consider these prespecified, method-specific tuning efforts broadly comparable in scope, while acknowledging that their search spaces are not identical. No holdout labels were used for selection.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised manuscript and point-by-point responses to all reviewers

Summary of changes in the revised manuscript

The method and the three-task results are unchanged. The revision strengthens positioning, evidence reporting, and ablation controls. Revision colors in the PDF mark provenance: red = hC7a, blue = 5RxN, green = X9pt.

Structural additions

  • §2 Related Work (green): contrasts with unrolled sparse coding, feature-level sensor–source concatenation, and normative deviation.
  • §3.3 (green): separates physiology-grounded quantities from pre-registered modeling choices and states what each pretraining objective preserves.
  • §3.4 (green): states the shared low-rank basis as a capacity control while subject coefficients and top-m retain individual variation.
  • §4.1/§4.2: cohort and baseline protocol, paired branch statistics, and complementarity reporting.
  • §4.4: pretraining four-condition table, K=1 both-off control, threshold diagnostics, rank sweep.
  • §5.1/§5.2/§6: fusion wording aligned to paired intervals, redundant restatement removed.
  • Appendix F: Tables 7–10; Figures 8–12 (reviewer-requested diagnostic figures).

Deletions and reasons

  1. §4.1 absolute subject-ID disjointness ("These corpora are separate ... is verified.") : replaced by a dataset/site-level claim, because public datasets do not document subject identity.
  2. §4.4 "Removing the healthy pretraining degrades ... helps the source branch." : replaced by the four-condition controlled table.
  3. §4.4 "Removing the multiscale temporal context degrades MDD ranking ... (Appendix F)." : deleted after verification: the numbers came from an earlier manuscript version/old protocol, are inconsistent with the final strict-holdout results, and no appendix table supported the sentence.
  4. "matches the best single (expert|branch) within seed noise" variants in §1/§4.2/§5.1/§5.2/§6 : replaced by paired-bootstrap intervals and a qualitative conclusion.
  5. §4.2 duplicate paired numbers and the trailing "+0.18" clause : removed; numbers now live in Table 12 and §4.2's first results paragraph.
  6. §4.1 "Table 2 ... lists all reproduction results ..." : folded into the revised baseline paragraphs.
  7. §4.2 repeated clause "The source branch is the strongest single expert on FEP (AUC 0.794), and" : removed; the PD 11-vs-6 evidence is kept.
  8. §4.2 parenthetical "source-minus-scalp penalty −0.08 to −0.17;" : removed as a duplicate of the preceding range.
  9. §6 separate sentence "It exceeds the source branch on PD." : merged into the preceding three-task clause.
  10. Appendix E, Table 6 "Head (with FL-DPD) MLP 42320→256→2" : relabelled as a development-only g_norm readout, because the reported Δh decision path does not use this head and it is excluded from the ≈169k parameter count.

Reruns

  • Scratch: random initialization with no pretrained checkpoint (the checkpoint path is left empty), keeping the same configuration, split, seed, and selection ("unified rerun"). The earlier scratch values came from a different configuration, so the rerun makes the four conditions strictly comparable. Reported in Table 8 and Figure 9.

New experiments added in the revision

  • K=1 both-operators-off control (rank-matched), Table 9 and Figure 8.
  • Extended rank sweep (additional ranks for MDD/FEP/PD, plus FEP per-coordinate), Table 10.
  • Threshold diagnostics, Figure 11.
  • Paired bootstrap intervals, Tables 8, 9, and 12 and Figure 10.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.