acceptodds
Under review as a conference paper at ICLR 2027

Shared Structure, Individual Dynamics: Flow Matching for EEG Generation

Abstract

Clinical electroencephalography (EEG) generation requires both patient-shared signal structure and sample-specific temporal detail. We study how the source distribution and training objective can address these requirements in Flow Matching. We propose GEMIA, a conditional Flow Matching method with a Global Structured Endpoint (GSE) and Multi-Scale Increment Alignment. GSE estimates spectral–spatial structure with equal weight for each training patient and models waveform shape and amplitude separately. The network predicts the clean endpoint while the transport loss is in velocity space; to limit endpoint error amplification, we cap the time-dependent weight in the velocity-\(L_1\) loss. Auxiliary time-domain losses complement the structured source by aligning per-channel scale, target temporal increments at several lags, and waveform correlation. We evaluate seven generators on four patient-wise TUH protocols, assessing signal fidelity, downstream utility, and memorization. GEMIA achieves the lowest mean TS-FID on all four protocols, with reductions of \(4.7%\)–\(63.1%\) relative to JET. It also achieves lower mean TS-FID with five Heun steps than the Flow Matching baselines with fifty steps. Source-objective ablations show that the benefit of GSE depends on the accompanying loss. These results support source-objective co-design for EEG generation.

Primary area:
generative models
Keywords:
EEG generation, Flow Matching, structured source distribution, multi-scale increment alignment, conditional generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_r8WmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GEMIA, a conditional Flow Matching method for EEG generation that combines a patient-balanced structured source with time-domain supervision. The source models cross-spectral structure and amplitude separately, while the objective combines capped velocity-L1 regression with channel-scale, multi-scale increment, and waveform-correlation losses. The method is evaluated against six generators on four patient-disjoint TUH protocols using fidelity, downstream utility, structural diagnostics, and sampling-step comparisons. GEMIA achieves the lowest mean TS-FID across the four protocols and maintains favorable TS-FID with fewer sampling steps.

Critical strengths
  1. The motivation is clear and the paper is easy to follow.
  2. The paper addresses the important problem of generating realistic multichannel EEG and proposes an interpretable division between population-level source statistics and sample-level supervision.
  3. GEMIA consistently achieves the lowest mean TS-FID among the compared methods across the four protocols.
Critical weaknesses
  1. All four evaluation protocols are drawn from the TUH corpus family and use a common 16-channel bipolar montage. Robustness to external datasets, alternative montages, or different acquisition devices is not evaluated. This limits the evidence for the method’s applicability across heterogeneous EEG recording settings. The authors should test this method on more datasets.
  2. Though the results is better, I hope the authors can provide more detailed ablation and comparison with JET. To me the novelty is not sufficient enough, as JET already provide a framework of flow matching plus EEG specific regularizations.
  3. Can the authors provide detailed downstream experiment settings? The TUEV balanced accuracy in Table B3 is specifically low.
  4. To better understand the generation quality, the authors can provide direct visualizations of generated signals and real signals. It can better show the transient events like spikes of generated signals.
  5. Additional comparisons of Hjorth parameters would also help assess temporal characteristics beyond spectral similarity.
Questions for authors

Please refer to the weakness.

3 comments
AuthorsOpenReview ↗

Response to r8Wm: external data, matched JET/DiT comparison, downstream protocol, and waveform/Hjorth analyses

We thank the reviewer for the careful reading and for the concrete requests.

1. External data (data breadth). For data breadth, four external datasets with the same Gaussian-vs-GSE control (same J3 objective, matched within dataset):

DatasetnGaussianGSEImprovement
SWEC-ETHZ iEEG3976.99 +/- 20.53638.51 +/- 297.44+34.6%
ISRUC-Sleep31809.62 +/- 13.241542.44 +/- 29.27+14.8%
HMC-Sleep31507.60 +/- 39.011218.93 +/- 133.03+19.1%
EEGMAT11707.031698.28+0.5%

iEEG is protocol-verified (train-only scaling; two seeds improve strongly, one ties). ISRUC/HMC are exploratory (amplitude scale fitted on all selected subjects); EEGMAT is a single-run control.

2. Novelty relative to JET. Contribution: patient-shared spectral structure in the source + an objective that exploits it. (i) GSE covariance root preserving within/cross-channel correlation vs isotropic Gaussian. (ii) Equal patient weight (each window 1/(P n_p)). (iii) Capped velocity-L1 bounds endpoint-error amplification near t=1. (iv) Auxiliary per-channel log-scale, multi-lag increment and waveform correlation vs the real target; not TV.

Three-seed ablation (TUSZ v1.5.2):

SourceCapped loss onlyFull objective
Gaussian1155.15 +/- 59.09448.01 +/- 37.07
GSE1120.07 +/- 106.78174.14 +/- 30.79

Source effect 3.04% under capped loss, 61.13% under the full objective, auxiliary-term effect 84.45% with GSE fixed. Across the four corpora, source effect under the full objective is 16.25-61.13%, cap-only -14.36% to +34.11%, and auxiliary effect 6.70% (TUAB) to 84.45% (TUSZ1.5.2). The source benefit is objective-dependent (the co-design claim). GEMIA has the lowest TS-FID among the four configurations in all 12 corpus-seed comparisons.

3. Matched JET/DiT comparison. Full source x objective ablation over three seeds on four corpora (per-seed table in supplement): JET and GEMIA use the same backbone, data loader and per-corpus budget, differing only in source and objective; DiT-FM is a same-protocol FM baseline with a DiT backbone. Three-seed TS-FID:

CorpusJETDiT-FMGEMIA
TUSZ v1.5.2472.24 +/- 18.92519.20 +/- 59.84174.14 +/- 30.79
TUSZ v2.0.6380.85 +/- 12.42511.54 +/- 11.59169.46 +/- 14.41
TUAB v3.0.1377.52 +/- 2.61424.97 +/- 19.04278.99 +/- 4.02
TUEV v2.0.1332.13 +/- 2.501053.42 +/- 48.91316.61 +/- 4.50

GEMIA beats JET and DiT-FM on all four corpora (JET paired reductions 63.2/55.6/26.1/4.7%); log-PSD favors GEMIA 4/4, ACF 2/4.

4. Downstream setting. TUEV: 6-class, 5-s windows, disjoint patient-wise splits; generator uses all training patients. Classifier real data = fixed 10% patient subset (seed 4407; 25 patients, 6,207 windows); synthetic added per class at synthetic:real 0.5; frozen CBraMod + linear head, AdamW, 50 epochs, batch 256, checkpoint by real-validation AUROC; three paired seeds (0/2025/2026). Test is held-out patients only; Table B3 is mean±sd over these runs.

5. TUEV BA (why it is low). Real-only 0.210; WGAN-GP 0.233; EEG-GAN 0.226; DDPM 0.222; MEG-GPT 0.209; JET 0.218; DiT 0.216; GEMIA 0.222. Test supports [567,4677,1998,329,2204,19646]: background dominates and rare classes are scarce in 5-s windows, so every classifier—including Real-only—predicts mostly the two majority classes (per-class recall behaves similarly). Low BA is a property of this low-patient protocol, not GEMIA-specific; we make no class-level claim. WGAN-GP/EEG-GAN have TS-FID 734.6/954.2 vs 316.6.

6. Waveforms (spike/transient) and Hjorth. We completed direct visual comparisons (real / GEMIA / JET / DiT) on four corpora: full-window raw waveforms (no crop/y-limit), a z-scored view, and zoomed panels around the largest-amplitude transient (fixed rule, not expert labels); real/generated are not pointwise pairs. In these windows GEMIA retains sharp spike-like transients; JET/DiT appear smoother. The rebuttal form does not support figure uploads, so the figure and window list will be in the camera-ready (available by email if the AC permits). Hjorth (single-seed; mobility/complexity, real/GEMIA/JET/DiT):

CorpusMobilityComplexity
TUSZ v1.5.20.331/0.392/0.357/0.7705.31/5.41/5.77/3.12
TUSZ v2.0.60.353/0.244/0.327/0.7554.83/8.29/6.29/3.05
TUAB v3.0.10.470/0.106/0.228/0.7234.46/16.35/9.35/3.75
TUEV v2.0.10.426/0.151/0.273/1.1924.91/12.55/7.34/1.77

TS-FID/log-PSD gains do not imply uniform time-domain gains: GEMIA complexity is higher than real on all corpora; we report this trade-off openly. The TUAB JET row uses the matching 28-epoch pool.

We appreciate your assessment of whether these controls address the concerns about novelty, ablation granularity, and the time-domain/Hjorth trade-offs.

I sincerely thank the authors for providing additional experiments and clarification. However, since the gains come mainly from the auxiliary losses rather than the structured source, my concern about novelty relative to JET remains, so I will keep my score.

AuthorsOpenReview ↗

Thank you to Reviewer r8Wm

Dear Reviewer r8Wm,

Thank you very much for your time and patience throughout the review and the follow-up discussion. Your suggestions on external data, a more detailed ablation and comparison with JET, the downstream setting, waveform visualizations, and Hjorth parameters led us to add several analyses and to understand the boundaries of our contribution more clearly. Your comments also helped us describe the relationship between the structured source and the auxiliary objective more precisely; we will make this distinction clearer in the revised manuscript and keep the trade-off discussion honest. We sincerely appreciate your rigorous review and your help. We wish you all the best.

— The GEMIA authors

Reviewer_Hn4GOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GEMIA, a conditional Flow-Matching method that improves the quality and efficiency of synthetic EEG data generation by leveraging a label-neutral, patient-balanced inductive bias. This promotes spectral distribution and cross-channel relationships across patients while maintaining desired sample-level properties. Comprehensive results are provided to evaluate the fidelity of generation, downstream impacts, and ablation.

Critical strengths
  1. Comprehensive evaluation of signal fidelity using various metrics, highlighting complementary properties.
  2. Downstream evaluation, ablation, and efficient generation results look promising and solid.
Critical weaknesses

The paper is relatively hard to follow fully at the beginning. The motive and objectives can be formulated more clearly without too many technical ties to flow matching (the introduction reads like related work).

Questions for authors
  1. The evaluation is based only on TUH, with mainly binary tasks except for a 6-class event. I would suggest including more datasets and more complex downstream tasks, e.g., those that have been used in "B[FM]2: Brain Foundation Model via Flow Matching with SplitUNet" (https://jd730.github.io/projects/BFM2/). A comparison with B[FM]2 would also be useful.

  2. It would also be great to see how fidelity changes across other (higher SNR) modalities like iEEG. Here are some large datasets: https://huggingface.co/datasets/NeuroTec/SWEC_iEEG_Dataset

1 comment
AuthorsOpenReview ↗

Response to Hn4G: writing revision, external iEEG/sleep evaluation, and SNR quantification

We thank the reviewer for the constructive review, and we agree the opening was too close to a related-work survey and buried the motivation. In the camera-ready we will lead with the two-part motivation (patient-shared structure vs. individual dynamics), state goals and contributions up front, and move flow-matching details to the method section.

Motivation. EEG generators are usually trained from a simple source, so the model must learn both the patient-shared spectral structure and the individual dynamics. We observe that the shared part is stable enough to estimate from training patients and move into the source; the remaining problem is the individual dynamics. The objective enables this: in a three-seed 2x2 ablation (TUSZ1.5.2), Gaussian+capped 1155.15, GSE+capped 1120.07, Gaussian+full 448.01, GSE+full 174.14. The source effect is 3.04% under capped loss and 61.13% under the full objective; adding the auxiliary terms with GSE fixed gives 84.45%. Across four corpora the source effect under the full objective is 16.25-61.13%, while cap-only ranges from -14.36% to +34.11%. We will rewrite the introduction around this objective-dependent benefit.

External evaluations. We evaluated four external datasets: iEEG (three training seeds), two 5-class sleep-staging corpora (three seeds each), and EEGMAT (single run). All controls use Gaussian source plus the same J3 objective, matched within each dataset.

DatasetnGaussianGSEImprovement
iEEG3976.99 +/- 20.53638.51 +/- 297.44+34.6%
ISRUC-Sleep31809.62 +/- 13.241542.44 +/- 29.27+14.8%
HMC-Sleep31507.60 +/- 39.011218.93 +/- 133.03+19.1%
EEGMAT11707.031698.28+0.5% (neutral)

iEEG (SWEC-ETHZ long-term; 8 patients, patient-wise 5/1/2, 16 contacts, 5-s at 200 Hz, train-only scaling) is protocol-verified; two seeds improve strongly and one ties the baseline. ISRUC-Sleep (40 subjects, 6 derivations, 30-s) and HMC-Sleep (24 subjects, 4 channels, 30-s) are exploratory under the current preprocessing protocol (amplitude scale estimated on all selected subjects); we report them as such. EEGMAT is a single-run neutral control.

SNR of the evaluated modalities. To make "higher-SNR" concrete, we computed a scale-invariant spectral SNR proxy on 2,000 test windows per corpus: 10 log10(P(1-45 Hz)/P(45-75 Hz)), median over windows/channels:

DatasetModalitySNR proxy (dB)
iEEG (SWEC-ETHZ)intracranial26.9 (23.6-31.4)
TUSZ v1.5.2scalp19.7 (14.2-24.4)
TUSZ v2.0.6scalp19.6 (12.5-25.0)
TUABscalp13.4 (8.9-17.9)
TUEVscalp15.6 (11.2-19.9)

iEEG is 7.3-13.2 dB higher than the four scalp corpora; on this higher-SNR modality the structured source still improves TS-FID by 34.6%.

Higher-SNR modality (iEEG). SWEC-ETHZ iEEG is the requested higher-SNR (intracranial) modality; we use its 18-subject long-term precursor (the larger SWEC iEEG release has 68 subjects and 4.6 TB, which we could not preprocess within the rebuttal window) with a patient-wise 5/1/2 split over 8 patients. We quantify fidelity change with the same strict-test evaluator: TS-FID improves from 976.99 +/- 20.53 to 638.51 +/- 297.44, a 34.6% reduction (three seeds; two improve strongly, one ties the baseline). Seed-0: log-PSD and ACF (0.5184 -> 0.1740) improve; alpha log-PSD and CSD worsen (single-seed).

Protocols. iEEG (SWEC-ETHZ long-term): 8 patients; patient-wise train ID01/ID08/ID13/ID16/ID18, validation ID11, test ID15/ID17; 16 contacts; 5-s windows at 200 Hz; band-pass 0.5-75 Hz; 18,720 training and 2,880 test windows (82 ictal); amplitude scale estimated from training windows only (median raw RMS 39.8369). ISRUC-Sleep: 40 subjects; patient-wise 30/5/5; 6 derivations; 30-s at 200 Hz; 35,766 windows. HMC-Sleep: 24 subjects; patient-wise 16/4/4; 4 channels; 30-s at 200 Hz; 22,608 windows. EEGMAT: 16 channels; 5-s; single run. Three training seeds (0/2025/2026) for the seeded datasets, sampling seed 0.

What these do and do not show. ISRUC and HMC are 5-class conditional-generation evaluations; we did not train a 5-class downstream classifier. The SNR above is a spectral proxy, not a calibrated amplifier SNR.

B[FM]2. We cannot run a fair comparison within the rebuttal window: it is a foundation-model pretraining approach (SplitUNet, ~6 M parameters, pretrained on ~307 h of TUEG, then fine-tuned per task), no public code/checkpoints are available for a matched comparison. We cite its reported conclusions and discuss the setting difference in the revision.

Scope. The four TUH protocols remain our controlled testbed; iEEG is protocol-verified, the sleep corpora are exploratory, and EEGMAT is a single-run control.

We appreciate your assessment of whether these evaluations address the concerns about scope, external validation and motivation.

Reviewer_fRaNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This article proposes GEMIA, a conditional flow matching method for multi-channel EEG generation. This method constructs a structured source distribution of patient balance through global structured endpoints, predicts clean endpoints using weighted truncated velocity L1 loss, and aligns channel scales, multiple time lag increments, and waveform correlations using multi-scale time assisted loss. The author evaluated the method on four TUH experimental protocols and compared it with six baseline methods.

Critical strengths

S1. The paper considers both the shared spectral structure and sample specific temporal dynamics in EEG generation.

S2. GSE is used to model shared structures, while time loss is used to model sample specific dynamics.

S3. The patient balance weighting method can avoid patients with a large number of windows dominating the training data.

S4. The paper includes four datasets, seven methods, ablation experiments, efficiency testing, and memory based diagnosis.

Critical weaknesses

W1. Both GSE and TS-FID use spectral information, so the performance improvement may partially come from feature overlap. Adding other independent indicators or quantitatively analyzing the degree of overlap between the two would be more convincing.

W2. The performance of non stream matching baselines is weak, making it difficult to explain the correlation. Comparing with JET for conditional matching is more meaningful.

W3. In Table 3, the contribution of GSE seems to be greater than the auxiliary loss. The author needs to provide a clearer explanation of the contributions of each component.

W4. Some supplementary indicators are not as good as JET. Including Amp-W1 and Env-W1 on TUSZ, as well as ACF on TUAB. The author should discuss the trade-offs between these indicators more clearly and provide absolute values, not just rankings.

W5. Supplementing training time, GPU usage, and inference latency can help evaluate the actual application cost of this method.

Questions for authors

Q1. To what extent do GSE and TS-FID overlap in spectral characteristics? Can quantitative analysis be provided?

Q2. Why is the result of GSE plus complete auxiliary loss in Table 3 lower than the three sub averages? Does this come from random seed differences?

Q3. When there is a significant difference between the number of patients and the number of windows, is the patient balance weighting still stable? Is a minimum number of patients required?

Q4. Can GEMIA and JET be compared under the same training budget and evaluation protocol?

4 comments
AuthorsOpenReview ↗

Response to fRaN: matched ablation and quantitative source–metric overlap analysis

Thank you for the detailed review.

W1 (source-metric relation). Direct source vs trained (TUSZ1.5.2/TUSZ2.0.6/TUAB/TUEV): GSE 298/337/270/420; Gaussian 1275-1364; GEMIA 151/161/280/318. Direct GSE is 4.3x closer than Gaussian and training reduces it further on 3/4 corpora (TUAB exception). GSE and TS-FID share spectral information (profile correlation 0.988-0.995); we do not claim to rule out overlap. Mechanism: the covariance-root source carries shared spectral-spatial second moments; auxiliary losses align the endpoint with the real target. log-PSD improves 4/4, ACF 2/4; amplitude/envelope worse.

W2 (baselines). We do not rely on non-FM baselines for the main claim. The core comparison is within one flow-matching framework, matched per corpus (TUSZ 200/256; TUEV 131/256; TUAB 28/128); JET and GEMIA use the same backbone, data loader and per-corpus training budget, differing only in source and objective (TUSZ v1.5.2: 7,578 vs 7,544 s). Three-seed TS-FID:

CorpusJETGEMIA
TUSZ v1.5.2472.24 +/- 18.92174.14 +/- 30.79
TUSZ v2.0.6380.85 +/- 12.42169.46 +/- 14.41
TUAB v3.0.1377.52 +/- 2.61278.99 +/- 4.02
TUEV v2.0.1332.13 +/- 2.50316.61 +/- 4.50

GEMIA also beats DiT-FM (519.2/511.5/425.0/1053.4 vs 174.1/169.5/279.0/316.6).

W3 (contribution attribution). Three-seed 2x2 on TUSZ v1.5.2:

ConfigurationTS-FID
Gaussian + capped loss1155.15 +/- 59.09
GSE + capped loss1120.07 +/- 106.78
Gaussian + full objective448.01 +/- 37.07
GSE + full objective (GEMIA)174.14 +/- 30.79

The source effect is 3.04% under the capped loss, 61.13% under the full objective, and adding auxiliary terms with GSE fixed gives 84.45%. Across corpora the source effect under the full objective is 16.25-61.13% (cap-only -14.36% to +34.11%); the auxiliary-term effect is 6.70% (TUAB) to 84.45% (TUSZ1.5.2). We claim objective-dependent benefit, not a fixed split. With GSE fixed, seed-0 progressive terms (TUSZ1.5.2/TUSZ2.0.6/TUAB/TUEV): none 1018.9/821.3/290.0/673.8; +scale 1237.2/1204.1/1584.1/379.4; +increment 456.5/325.2/312.9/463.8; +all 150.9/161.5/279.6/317.9. The objective side is not a small add-on: with either source it lowers TS-FID by 61.2% (Gaussian) and 84.5% (GSE) in 3-seed means. Scale alone is mixed, increment helps 3/4, correlation improves 4/4.

W4 (metric trade-offs). GEMIA-minus-JET: log-PSD -0.216/-0.798/-0.676/-0.349; ACF -0.048/-0.015/+0.075/+0.049; Amp-W1 +0.533/+0.389/+0.091/+0.055; Env-W1 +0.853/+0.611/+0.148/+0.081. The rank map is only a visualization; absolute values are in Appendix Table B2.

W5 (cost). Training: 7,544/8,758/8,793/15,982 s. 129.86 M parameters; K=5 gives 9 evaluations; 5.322/5.291/5.439/8.791 s per 1,000 samples.

Q1 (quantitative overlap). TS-FID decomposes into a mean-shift and a covariance term: ||mu - mu_r||^2 (mean) and tr(Sigma + Sigma_r - 2(Sigma Sigma_r)^1/2) (covariance). TUSZ v1.5.2 (20,160 test windows):

ConfigurationmeancovarianceTS-FID
Gaussian source709.3565.91275.2
GSE source142.1156.0298.1
GSE + capped loss831.0187.91018.9
GEMIA48.2102.8150.9

GSE cuts the mean term by 80.0% and covariance by 72.4% relative to Gaussian - this is the structural overlap. But source alone is 298 vs the real-real floor 7.8; GEMIA reaches 150.9, and its residual is 68.1% covariance. With GSE fixed, the objective cuts mean 831.0->48.2 (94.2%) and covariance 187.9->102.8 (45.3%). Overlap is real but does not explain the trained gain.

Q2 (three-seed ablation table). The original table used seed 0; per-seed TS-FID:

Corpusseed 0seed 2025seed 2026mean +/- sd
TUSZ v1.5.2150.92209.07162.43174.14 +/- 30.79
TUSZ v2.0.6161.49186.09160.80169.46 +/- 14.41
TUAB v3.0.1279.62282.67274.70278.99 +/- 4.02
TUEV v2.0.1317.87320.35311.62316.61 +/- 4.50

The complete configuration wins every tested seed.

Q3. Each patient contributes equal total weight, with each window weighted by 1/(P n_p), targeting patient-average spectral statistics. Independent patient-level estimates with bounded variance give O_p(P^{-1/2}) sampling error; within-patient correlation is absorbed in patient-level variance. Equal weighting prevents long recordings from dominating; the required patient count depends on heterogeneity, noise and precision; no universal minimum.

Q4. Yes: within each corpus JET and GEMIA share split, loader, batch size, learning rate, epochs and backbone (129.86 M); all JET numbers are our re-runs, not quoted. TUSZ v1.5.2: 7,578 vs 7,544 s; paired reductions 63.2/55.6/26.1/4.7%. Per-seed values for these new three-seed runs are in the supplementary table for this response (the original submission's ablation table remains seed-0).

Thank you for the authors' detailed response and supplementary experiments. The comparison with JET under the same training budget, the 2 × 2 ablation experiments with three random seeds, and the computational cost data respond to the corresponding points. But my core concern (W1/Q1) has not been resolved yet, and the newly provided numbers make this issue even more prominent: 

  • According to the authors' response to W1, samples were taken directly from GSE without any network, and the TS-FID values were 298/337/270/420, respectively. This is lower than the trained JET on three of the four datasets, lower than DiT-FM on all four datasets, and even lower than GEMIA itself on TUAB (270 vs 279). Please add a row to Table 2 that reads "GSE directly sampled, untrained" as a baseline. 
  • In Table B2, GEMIA performs worse than JET on RFD on all four datasets, as well as on Amp-W1 and Env-W1; GEMIA's BP-FD is about 16 times worse than JET's on TUSZ v1.5.2. In Table B4, CSD error is the most direct test for the cross-spectral structure that GSE should carry, and GEMIA's is worse than JET's on TUSZ v1.5.2 (2.32 vs 0.99, the highest among all generators) and TUEV. The authors' response to W4 did not list RFD and BP-FD. May I ask how these results can be reconciled with the statement of 'achieving state-of-the-art fidelity'? 
  • In the authors' Q1 decomposition, training only with L_cap increases TS-FID from 298 to 1019, with the mean term increasing from 142 to 831. That is to say, under the transport loss, the network actually deteriorates its source distribution, and the so-called 84% auxiliary loss contribution is measured relative to this already deteriorated baseline. What is the reason for this? 
  • In Table B3, after adding GEMIA-generated data on TUSZ v1.5.2, TUSZ v2.0.6, and TUAB, at least two of the three metrics are lower than the results obtained using only real data; all three metrics improve only on TUEV.. Given that the motivation claimed in the paper is data augmentation in low-resource scenarios, this should be clearly stated in the main text.
AuthorsOpenReview ↗

Follow-up: direct-GSE baseline, full fidelity ledger, and augmentation outcomes

We thank the reviewer for recognizing the matched-budget JET comparison, three-seed factorial ablation, and computational-cost analysis. We will make the direct-source baseline, metric-specific conclusions, and augmentation outcomes explicit in the main paper.

1. Direct-GSE baseline. We will add "GSE directly sampled, untrained" to Table 2, with TS-FID 298.12 / 336.67 / 270.35 / 419.68 (TUSZ v1.5.2 / TUSZ v2.0.6 / TUAB / TUEV). This is a valid stochastic sampling baseline using statistics estimated from training patients. Its current implementation is label-neutral, while GEMIA is class-conditioned; this distinction does not diminish its relevance to distributional evaluation. Its strong scores demonstrate that the source already captures substantial structure rewarded by TS-FID. On TUAB, direct GSE outperforms GEMIA (270.35 vs 279.62, seed 0). On the class-stratified audit subset used for signal-level diagnostics, direct GSE also gives log-PSD 1.44/2.37/1.81/2.70, ACF 0.038/0.055/0.026/0.056, Amp-W1 0.446/0.553/0.478/0.400, Env-W1 0.677/0.839/0.720/0.629, CSD 1.015/0.972/0.984/4.587 and BP-FD 196/81/156/69,351 (TUSZ v1.5.2 / TUSZ v2.0.6 / TUAB / TUEV). These diagnostics extend the direct-source evaluation beyond TS-FID. We report the label-neutral source alongside the class-conditioned model under the same diagnostic protocol. We will explicitly distinguish gains over direct sampling from gains over trained comparators and include the direct-source baseline in the applicable complementary evaluations.

2. Complete fidelity results. We will remove the unqualified "state-of-the-art fidelity" claim. The supported positive result is lower TS-FID and log-PSD error than matched JET on all four corpora. RFD, Amp-W1, and Env-W1 are worse on all four; BP-FD is worse on both TUSZ versions and better on TUAB/TUEV; CSD error is worse on TUSZ v1.5.2 (2.32 vs 0.99) and TUEV. W4 should have included these results, particularly the large BP-FD gap on TUSZ v1.5.2. We will discuss them in the main text alongside the improvements. Since BP-FD and CSD also assess spectral properties, the revised claims will name the improved metrics explicitly. Encoding spectral-spatial moments in the source does not guarantee their preservation in the learned endpoint distribution.

3. Why the capped-loss configuration deteriorates. On TUSZ v1.5.2, the seed-0 comparison is direct GSE 298.12, GSE + L_cap 1018.89, and GEMIA 150.92. The deterioration is dominated by the TS-FID feature-mean term (approximately 142 to 831), while the feature-covariance term changes from 156 to 188. These are moments in the evaluation feature space, not the raw-signal cross-spectral covariance.

Theoretically, coordinatewise L1 regression targets conditional medians, whereas standard squared-L2 flow matching targets the conditional-mean velocity. The standard marginal-path matching guarantee therefore does not automatically transfer to L_cap. Time capping reweights the regression loss; it does not constrain final-sample TS-FID moments. This explains why deterioration is possible, although the present ablations do not isolate its cause. The auxiliary terms supervise endpoint scale, target-matched increments, and waveform correlation; empirically, the full model reduces both TS-FID components (to approximately 48 and 103).

We will report the reference points separately: the seed-0 reduction is 85.2% relative to GSE + L_cap and 49.4% relative to direct GSE. The previously reported 84.45% compares three-seed means against the capped-loss configuration; it is not a percentage allocation of contribution.

The source benefit is also supported independently of that deteriorated reference: with the full objective fixed, Gaussian + full scores 448.01 +/- 37.07 and GSE + full 174.14 +/- 30.79 on TUSZ v1.5.2, a 61.13% reduction. Under the capped loss, the corresponding three-seed source effect is 3.04%, showing its dependence on the objective. Across all four corpora, the full-objective source substitution improves the three-seed means by 16.25-61.13%. These matched comparisons support an objective-dependent source benefit, while the direct-source comparison quantifies the additional effect of training.

4. Augmentation outcomes. We will state in the main text that, under the evaluated protocol, augmentation lowers at least two of BA/AUPRC/AUROC on both TUSZ versions and TUAB. In particular, BA decreases from 0.692 to 0.584 on TUSZ v1.5.2 and from 0.713 to 0.653 on v2.0.6. All three improve only on TUEV: 0.210/0.266/0.721 to 0.222/0.282/0.781. The 10% patient restriction applies to the classifier's real training subset; the generator uses the full training-patient set. We will present augmentation as an evaluated application with mixed outcomes, rather than a demonstrated general low-resource benefit, and align the abstract, introduction, and conclusion with these results.

AuthorsOpenReview ↗

Thank you to Reviewer fRaN

Dear Reviewer fRaN,

Thank you very much for your detailed and demanding review, and for taking the time to follow up on our response. Your insistence on quantifying the overlap between the GSE source and the TS-FID feature space, adding a direct-GSE baseline, and reporting the complete fidelity ledger—including RFD, BP-FD and CSD—substantially improved the rigor of our analysis and made the boundaries of our claims much clearer. The 2×2 ablation, the mean/covariance decomposition, and the revised metric discussion in our responses are direct results of your guidance. We are very grateful for the care and expertise you brought to this review, and we will carry these corrections into the revision so that the paper's claims match its evidence. We wish you all the best.

— The GEMIA authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.