{"id":"a7254bd9-5cc9-4b7f-9c2c-e9fe23156ba5","arxiv_id":"2506.02433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A pre-trained diffusion-based generative framework synthesizes fMRI BOLD and fNIRS signals from EEG alone, with moderate consistency and improved fairness after data augmentation.","lead":"A machine-learning framework is proposed to turn cheap EEG brain recordings into synthetic fMRI or fNIRS scans, which could make advanced brain imaging more affordable. The authors report that the generated images match real recordings at moderate levels and that augmenting underrepresented classes with the generated data improves fairness in brain-computer interface decoding.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fairness claim rests on a circular evaluation: augmented training rows only help if the classifier treats generated fMRI as the same task-conditioned distribution as real data, and the paper never validates this against simpler non-generative baselines.","rationale":"The reader's verdict is CONDITIONAL with a medium correctness risk, focusing on small-sample generalization. I agree that sample size is a serious concern, but the single most load-bearing issue in the argument is the fairness augmentation claim. This claim is the abstract's headline application ('enhance model fairness by augmenting data for underrepresented groups') and the quantitative evidence in Table 1 is the strongest single result in the paper. Yet the experimental design does not establish that the generative model causes the improvement. The only comparison is before vs. after augmentation with generated samples; there is no comparison to random oversampling or class-weight reweighting, which are standard, cheap baselines that mechanically increase minority-class recall and F1 in imbalanced classification. Because the model is trained on the same EEG inputs and evaluated with five-fold/three-fold cross-validation on the same data, the improvement could be entirely a result of the classifier seeing more minority-class examples, not of the generated examples being informative. The paper's own Limitations section and Extended Data Fig. 1 admit imperfect generation with regional and temporal mismatch, so we cannot assume the generated rows are class-conditional draws from the real distribution. Additionally, the absence of comparisons against the authors' own prior generative methods (CATD, SCDM in refs 30-31) weakens the novelty attribution of the unified representation module. This is not an internal inconsistency in the mathematics; it is an under-identification of the causal claim. The proposed concrete test is feasible because the datasets are open-access and the code, while not released, could be reconstructed from the methods; the controls require only the downstream classifier and standard data-processing routines. I would keep the CONDITIONAL verdict but tighten the condition: release code or results for the controls. As a stress-test reviewer, I am not arguing the method is fraudulent; I am arguing that the fairness benefit, as presented, is not yet attributable to the generative framework rather than to generic imbalance handling.","tokens_in":27403,"tokens_out":1849,"duration_ms":16579,"concrete_test":"Re-run the Table 1 fairness protocol with three controls on the same downstream classifier: (1) class reweighting of the original 30-sample minority class, (2) random oversampling with replacement to 150 samples without any generative model, and (3) the authors' own prior EEG-to-fMRI/fNIRS generative models (CATD, SCDM). If the proposed framework's minority-class F1 is not significantly above the best control (e.g., not above random oversampling), then the fairness improvement is not attributable to the unified generative representation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's most consequential claim is that augmenting underrepresented classes with generated BOLD/fNIRS improves BCI fairness. In Table 1, augmentation raises minority-class F1 from 0.539±0.397 to 0.881±0.040 (fMRI Task 5) and from 0.037±0.052 to 0.743±0.016 (fNIRS LMI). The load-bearing premise is that these generated samples carry task-discriminative signal aligned with real class distributions. However, the experiment never tests this premise against the simplest alternative: training the same downstream classifier on the original 30 minority samples plus 120 random copies, or with standard class reweighting/oversampling. If those baselines yield similar F1 gains, then the improvement is explained by generic imbalance correction, not by the proposed generative unified representation. Moreover, because the same encoder-extractor-generative pipeline supplies the augmented samples and the paper reports no comparison with the authors' own prior EEG-to-fMRI/fNIRS generative baselines (refs 30, 31, e.g., CATD and SCDM), the attribution of the fairness gain to the unified representation module is unsecured. The paper's own Extended Data Fig. 1 admits failures with regional and temporal mismatch in generated BOLD, so the generated samples are imperfect by the authors' account; without a non-generative control, the burden of proof for the fairness improvement falls on a circular demonstration that stronger class balance alone cannot explain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pre-trained generative framework that maps EEG into fMRI-like BOLD and fNIRS signals using pretrained feature extractors, a hyperdimensional integration module, and a Diffusion Transformer-based unified representation module. The authors evaluate the framework on EEG-fMRI and EEG-fNIRS datasets, reporting moderate temporal correlation (PCC 0.43–0.50 vs. noise 0.03–0.06), spatial similarity (SSIM approximately 0.81), cross-modal and cross-subject generalization, improvements in downstream decoding and clinical decision-support tasks, and fairness gains through data augmentation for underrepresented classes. The main claims are that the framework generates data consistent with real brain activity, provides interpretable brain-mechanism insights, lowers neuroimaging cost, and enhances fairness of BCI decoding models.","tokens_in":27745,"tokens_out":6465,"duration_ms":63958,"significance":"If the central claims hold, the framework would be a practically useful contribution to low-cost neuroimaging and BCI fairness: it provides a concrete pipeline from EEG to hemodynamic modalities, reports quantitative fidelity metrics against noise baselines, and demonstrates a plausible use of synthetic data for class rebalancing. The use of held-out test subjects and multiple datasets is a strength, as is the explicit comparison against a noise baseline for the generation fidelity metrics. However, the current evidence is insufficient to support several headline claims, particularly the fairness contribution, the 90% cost reduction, and the neurobiological interpretability of the SHAP analyses. The novelty relative to the authors' own prior CATD and SCDM methods is also not demonstrated, since no direct comparison is provided.","major_comments":[{"comment":"The fairness evaluation lacks any non-generative control. The reported improvements (e.g., Task 5 F1 from 0.539±0.397 to 0.881±0.040; LMI from 0.037±0.052 to 0.743±0.016) are compared only against the imbalanced baseline. Without comparing with simple oversampling of the original 30 minority samples, class reweighting, or SMOTE, the gains could be fully explained by generic class-balance correction, and the specific contribution of the proposed unified representation and generative module is not established. Furthermore, no comparison is made with the authors' own prior EEG-to-fMRI/fNIRS models (refs 30 and 31, CATD and SCDM), so the advantage of the proposed framework over existing unified-representation approaches is unverified. Because fairness enhancement is a central claim of the abstract and the Results, these control experiments are necessary before the claim can be accepted.","section":"Cross-Modal Data Augmentation Enhances Fairness of BCI Decoding Models (Table 1)"},{"comment":"The statement that the framework can 'reduce operational costs by an estimated 90%' is an unsupported quantitative claim; no cost model, calculation, or citation is provided. This figure appears in both the Results and the Discussion, and it substantially overstates what the current experiments can demonstrate. The abstract's use of 'accurately generates' is also stronger than the evidence supports: the PCC values of 0.43–0.50 and SSIM of 0.81 indicate moderate fidelity, and the paper's own Extended Data Fig. 1 documents regional and temporal mismatches. The cost-reduction figure should be removed or replaced with a derived estimate, and the accuracy language should be tempered.","section":"Results, paragraph after Fig. 5"},{"comment":"The SHAP analyses are presented as evidence that the framework 'interprets brain mechanisms' and 'captures biologically plausible representations,' but SHAP values on the model's own inputs and outputs measure the sensitivity of the learned mapping, not neurobiological causality. The overlap between SHAP-identified regions of real and generated BOLD and the dominance of the gamma band are consistent with known phenomena, but they do not by themselves establish that the framework provides mechanistic insight. For example, the gamma-band result could arise from the EEG feature extractor's encoding rather than from a true neurovascular coupling learned by the model. The authors should either validate these SHAP-based findings against independent neuroimaging evidence (e.g., known task-evoked networks or prior fMRI studies) or substantially qualify the mechanistic language used in the text and figure captions.","section":"The Unified Representations Provide Insights into Brain Mechanisms (Fig. 3)"},{"comment":"The cross-subject generalization claim rests on only three held-out subjects in Fig. 4(c,d), with no confidence intervals or statistical test across subjects, while the training sets contain 17–29 subjects. The paper's own Limitations section admits that 'most available datasets include fewer than 30 subjects' and that the model suffers from 'overfitting to dominant patterns' (Extended Data Fig. 1). Given this, the claim of 'strong generalization across modalities and subjects' is not supported by the scale or statistical treatment of the evidence. The authors should either add more subjects and report per-subject variability with appropriate statistics, or moderate the claim to something like 'preliminary evidence of cross-subject consistency on a small held-out set.'","section":"The Proposed Framework Enables Cross-Modal and Cross-Subject Generalization (Fig. 4(c,d))"}],"minor_comments":[{"comment":"The caption contains a typo: 'Simliar' should be 'Similar.'","section":"Fig. 5 caption"},{"comment":"The caption contains a typo: 'revelance' should be 'relevance.'","section":"Fig. 4 caption"},{"comment":"The sentence 'The Methods section is organized as follows: First, ... Next, ... Finally, ... This framework provides a comprehensive exposition of the framework, from data preparation to its operationalization and application.' is repetitive and should be rewritten for clarity.","section":"Methods, paragraph on organization"},{"comment":"The physiological delay τ is fixed to approximately 6 seconds both in the preprocessing (BOLD shifted by 6 s) and in Eq. (3). Please clarify whether τ is a fixed hyperparameter in all experiments and whether the reported PCC values are computed after the same shift; otherwise the temporal-consistency result may be partly an artifact of the alignment procedure.","section":"Eq. (3) and Preprocessing"},{"comment":"The 'noise baseline' used for PCC comparisons is never defined in the Methods. Please specify how the noise signals were generated (e.g., Gaussian white noise, phase-randomized surrogate, or shuffled real signals), as the validity of the noise comparison depends on this choice.","section":"Methods, Preprocessing and Fig. 2"},{"comment":"The abstract and Fig. 1 mention ECoG and fPAI as modalities within the unified representation space, but the experiments cover only EEG, fMRI, and fNIRS. Please clarify explicitly that ECoG and fPAI generation are future directions, not results of this work.","section":"General scope"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a strong central idea but currently overclaims on several fronts. The most important missing experiments are the non-generative baselines for the fairness table (simple oversampling, reweighting) and a direct comparison with the authors' own CATD and SCDM methods. If those baselines show that simple rebalancing produces similar F1 gains, the fairness claim should be dropped or reframed as 'data augmentation can improve fairness' without attributing the effect to the generative model. I also note the high proportion of self-citations to the authors' patents and prior work (refs 13, 14, 17–21, 27–31); this should be checked by the editor for whether it distorts the novelty narrative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the one-line take: the generation results are credible for this hard problem, but the fairness claim is not yet earned. The paper shows a unified framework that maps EEG to both fMRI and fNIRS, with held-out PCC of 0.43–0.50 versus noise 0.03–0.06 and SSIM 0.81. That is real signal, not just fitting training targets.\n\nWhat's new: the joint framework covering both modality pairs, the cross-dataset transfer (model trained on EEG-fMRI applied to EEG-fNIRS subjects), and the fairness augmentation protocol. The components—LaBraM features, DiT diffusion, hyperdimensional alignment—come from the authors' prior papers, but the combination and the cross-modal transfer experiment are not.\n\nThe soft spot is exactly where the stress-test note points. The fairness table shows large F1 gains for underrepresented classes (0.037 to 0.743 for fNIRS LMI), but there is no baseline of simple oversampling, class reweighting, or random duplication of minority samples. Without that control, the improvement may be generic imbalance correction, not the generative framework. The authors' own prior SCDM and CATD methods should also be baselines; their absence makes it hard to see the added value. This is the load-bearing weakness.\n\nOther issues are minor but real: the 90% cost reduction is an unbacked estimate, the abstract overuses 'accurate,' and the SHAP-based 'brain mechanisms' section is interpretive. The small datasets (17–29 subjects) and the authors' own Extended Data Fig. 1 showing regional mismatch mean the generalization claims should be read with caution. To their credit, the limitations section is candid about overfitting, which is more than most papers do.\n\nI'd accept this for review. The core generation evidence is strong enough to merit referee time, and the fairness control is a fixable experimental gap, not a fatal flaw. I'd want code, the non-generative baselines, SCDM/CATD comparisons, and error bars on the main PCC/SSIM statistics. A revised version could be a solid contribution.\n\nFor the reading group: maybe. It's a useful data point for anyone working on cross-modal neuroimaging generation, but the missing baselines make it more of a discussion item than a settled result.","headline":"A credible EEG-to-fMRI/fNIRS generation study with one unearned claim: the fairness gains are not benchmarked against trivial oversampling.","tokens_in":28244,"tokens_out":2702,"would_cite":false,"duration_ms":26077,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A unified generative model turns cheap EEG into fMRI-like BOLD signals","keywords":["functional neuroimaging","generative AI","unified representation","EEG-to-fMRI synthesis","diffusion model","brain-computer interface","model fairness","cross-modal generation"],"falsifier":"The deciding experiment is a cross-site transfer test: train the framework on simultaneous EEG-fMRI data from one scanner or site, generate BOLD for EEG from a second site or scanner with different acquisition parameters, and compare against real BOLD; if the cross-subject Pearson correlation falls to the noise baseline (about 0.03–0.06) or the SSIM drops far below 0.81, the claimed cross-subject and cross-task generalization fails. A complementary check is to apply the augmentation procedure to a dataset with genuine demographic imbalance and see whether the minority-class F1 gain persists.","tokens_in":27194,"feed_emoji":"🧠","tokens_out":9478,"duration_ms":77188,"temperature":0.7,"pith_summary":"Functional neuroimaging that reveals brain activity in detail—fMRI and fNIRS—is expensive, immobile, and hard to acquire, while EEG is cheap and ubiquitous. This paper argues that a pre-trained generative framework can learn a single representation space shared by EEG, fMRI, and fNIRS, and then reconstruct the expensive modality from the cheap one: from EEG alone it generates blood-oxygen-level-dependent (BOLD) time courses that correlate with real BOLD (Pearson correlation 0.43–0.50 versus a noise baseline near 0.03–0.06) and match its spatial structure at a structural similarity score (SSIM) of about 0.81. The same generated data, used as augmentation for an artificially imbalanced dataset, raises the minority-class F1 score from 0.539 ± 0.397 to 0.881 ± 0.040 in an fMRI visual-decoding task and from 0.037 ± 0.052 to 0.743 ± 0.016 in an fNIRS motor-imagery task. If these results hold, advanced functional neuroimaging becomes dramatically cheaper—the authors estimate a 90% reduction in operational cost—and brain-computer interface (BCI) decoding models trained on imbalanced brain data become fairer.","feed_headline":"Cheap EEG generates fMRI-like BOLD signals","feed_subtitle":"A unified generative model maps EEG, fMRI, and fNIRS into one space, then rebuilds the expensive signals and boosts fairness.","key_machinery":"The unifying mechanism is the 'unified representation space.' It is produced by three coupled components: pre-trained feature extractors for each modality; a hyperdimensional integration step that aligns the two modalities in space and time (electrode signals are re-weighted onto cortical sampling points by inverse squared distance, and time series are aligned by a Gaussian kernel matrix centered at the approximately 6-second hemodynamic delay); and a Diffusion Transformer that learns the joint distribution of the aligned representations. A modal-unpatcher decoder then reconstructs the target modality from a sample of the unified representation. This machinery is what transfers the temporal richness of EEG into the spatial detail of BOLD or fNIRS while keeping the physiological relationship between electrical and hemodynamic signals intact.","core_discovery":"On its own terms, the paper establishes that a single pre-trained generative pipeline—pre-trained feature extractors, a hyperdimensional spatial and temporal alignment step, a Diffusion Transformer generative module, and a modal-unpatcher decoder—can map simultaneously recorded EEG and either fMRI or fNIRS into one unified representation, and can then decode that representation back into the target modality. Trained on paired EEG–fMRI data, the model generates BOLD from EEG alone; the generated BOLD reproduce regional activations, functional-connectivity structure, and the known approximately 6-second hemodynamic delay. Trained on EEG–fNIRS data, it generates fNIRS oxy- and deoxy-hemoglobin signals from EEG. The authors further claim that using these generated samples to rebalance underrepresented classes reduces the performance gap between minority and majority groups, and that the unified representations support downstream decoding—visual stimulus, motor imagery, body-mass-index and age regression—at levels close to or beyond what real BOLD provides.","pith_inferences":["A natural reverse test the paper does not run: if the unified representation space is truly shared, the framework should also reconstruct EEG from fMRI or fNIRS; the paper only demonstrates cheap-to-expensive synthesis, so a symmetric experiment would probe whether the space is genuinely modality-invariant.","The fairness improvement is demonstrated on simulated task-level imbalance within small datasets, not on demographic imbalance; applying the same augmentation to a dataset with real group underrepresentation (by age, sex, or clinical status) would show whether the fairness claim extends beyond the simulated setting.","The temporal correlations reported (0.43–0.50) are modest in absolute terms, which suggests the framework's main added value may be spatial and structural fidelity (SSIM of about 0.81 and connectivity preservation) rather than high-fidelity time-course reconstruction; a comparison against a simple delayed linear regression on the same surfaces would settle this.","The paper's own limitation note—that most paired datasets have fewer than 30 subjects—makes cross-site generalization the most likely failure point; a held-out-site or cross-scanner evaluation would be the strongest test of the framework's practical promise."],"forward_implications":["fMRI and fNIRS signals could be synthesized from EEG alone in settings where only EEG recording is feasible, removing the primary cost and mobility barriers to multimodal functional neuroimaging.","Balancing an imbalanced BCI training set with generated samples should raise the performance of underrepresented classes, reduce prediction variance, and shrink the accuracy gap with well-represented classes.","Downstream decoding—visual stimulus identity, motor imagery state, and continuous physiological prediction—should reach levels close to those obtained with real BOLD, and in some cases (for example BMI prediction) surpass them.","The generated signals carry interpretable neurophysiological structure, including the approximately 6-second hemodynamic lag, gamma-band dominance in visual tasks, and functional connectivity patterns overlapping with real BOLD."],"supporting_citations":[{"why":"Supplies the Diffusion Transformer generative backbone that learns the joint distribution of aligned EEG and target-modality representations.","marker":"[52]"},{"why":"Open naturalistic-viewing dataset with simultaneous EEG and fMRI, used for within-subject and cross-subject BOLD generation and decoding tasks.","marker":"[53]"},{"why":"Resting-state simultaneous EEG-fMRI dataset used to train the EEG-to-BOLD model later applied to the Parkinson's EEG data.","marker":"[54]"},{"why":"Companion description of the same resting-state EEG-fMRI dataset and its preprocessing, used with [54] for the training set.","marker":"[55]"},{"why":"Open EEG-fNIRS motor-imagery dataset used to train and test fNIRS generation from EEG and the modality's fairness experiments.","marker":"[57]"},{"why":"Establishes the physiological approximately 6-second hemodynamic delay that the temporal alignment matrix is centered on.","marker":"[60]"},{"why":"Pre-trained EEG feature extractor whose parameters initialize the EEG encoder, providing hierarchical temporal and spatial patterns.","marker":"[61]"},{"why":"Large-cohort resting-state and task fMRI data used to pre-train the fMRI feature extractor in the temporal and spatial consistency experiments.","marker":"[62]"}],"fun_headline_variants":["Generative model turns EEG into fMRI-like BOLD signals","Unified generative framework boosts BCI fairness","EEG-to-fMRI generation improves BCI decoding fairness","Pre-trained generative model creates missing neuroimaging data","One generative model unifies EEG, fMRI, and fNIRS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a mapping learned from small paired datasets of roughly 17–29 subjects captures enough population variability that the same EEG-to-BOLD and EEG-to-fNIRS mapping works across new subjects, tasks, and clinical groups; the paper's own limitations acknowledge that most available datasets have fewer than 30 subjects and that overfitting to dominant patterns occurs.","fun_headline_variants_meta":{"raw":{"variants":["Generative model turns EEG into fMRI-like BOLD signals","Unified generative framework boosts BCI fairness","EEG-to-fMRI generation improves BCI decoding fairness","Pre-trained generative model creates missing neuroimaging data","One generative model unifies EEG, fMRI, and fNIRS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000899,"raw_usage":{"total_tokens":3850,"prompt_tokens":902,"completion_tokens":2948,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2880}},"tokens_in":518,"tokens_out":2948,"duration_ms":20890,"temperature":1.0,"reasoning_tokens":2880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:24:30.373142+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The deciding experiment is a cross-site transfer test: train the framework on simultaneous EEG-fMRI data from one scanner or site, generate BOLD for EEG from a second site or scanner with different acquisition parameters, and compare against real BOLD; if the cross-subject Pearson correlation falls to the noise baseline (about 0.03–0.06) or the SSIM drops far below 0.81, the claimed cross-subject and cross-task generalization fails. A complementary check is to apply the augmentation procedure to a dataset with genuine demographic imbalance and see whether the minority-class F1 gain persists.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Diffusion Transformer generative backbone that learns the joint distribution of aligned EEG and target-modality representations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Open naturalistic-viewing dataset with simultaneous EEG and fMRI, used for within-subject and cross-subject BOLD generation and decoding tasks."},{"cited_title":"W., Zhang, G","cited_arxiv_id":null,"evidence_quote":"Resting-state simultaneous EEG-fMRI dataset used to train the EEG-to-BOLD model later applied to the Parkinson's EEG data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Companion description of the same resting-state EEG-fMRI dataset and its preprocessing, used with [54] for the training set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Open EEG-fNIRS motor-imagery dataset used to train and test fNIRS generation from EEG and the modality's fairness experiments."},{"cited_title":"K., Pauls, J., Augath, M., Trinath, T","cited_arxiv_id":null,"evidence_quote":"Establishes the physiological approximately 6-second hemodynamic delay that the temporal alignment matrix is centered on."},{"cited_title":"& liang Lu, B","cited_arxiv_id":null,"evidence_quote":"Pre-trained EEG feature extractor whose parameters initialize the EEG encoder, providing hierarchical temporal and spatial patterns."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Large-cohort resting-state and task fMRI data used to pre-train the fMRI feature extractor in the temporal and spatial consistency experiments."}],"review_version":1}