{"id":"9f737717-ae6a-43b9-9af7-8ebb20140759","arxiv_id":"2509.10491","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"FlowECG shows that a flow matching version of SSSD-ECG generates 12-lead ECGs with quality comparable to the diffusion baseline while using only 10 to 25 sampling steps.","lead":"This paper replaces the slow diffusion process in an existing ECG generator with flow matching, a faster way to turn noise into heart signals. The authors report similar signal quality with 10 to 25 sampling steps instead of 200, which could make synthetic ECGs more practical in hospitals.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central efficiency claim rests on an unquantified NFE sweep: no numeric table, error bars, seeds, or statistical comparison support 'comparable at 10–25 NFE'.","rationale":"I read the paper in good faith. Flow matching is a standard, credible technique; replacing diffusion with flow matching in an SSSD-ECG-like architecture is a plausible engineering contribution. The 200-NFE metrics in Table 1 are favorable and not internally contradictory. The central risk is not the method but the support for the headline efficiency claim: it rests on a single figure with no numerical values, no variance, and no baseline validation. This is a concrete, testable weakness, not a fatal flaw. The reader's weakest-assumption statement already pointed at the lack of error bars/seeds and the retrained baseline; my stress-test sharpens this to the specific missing NFE-sweep evidence. Because the paper is already CONDITIONAL and my concern is the same kind of missing-evidence issue rather than an internal inconsistency, I recommend no change to the verdict. A numeric, seeded NFE sweep and baseline verification would settle the concern and could move the paper toward acceptance.","tokens_in":6662,"tokens_out":4439,"duration_ms":57097,"concrete_test":"Publish a numeric NFE-sweep table for both models over NFE ∈ {2, 5, 10, 15, 20, 25, 50, 100, 200}, reporting all four metrics as mean ± std across at least 3 independently trained seeds on the same held-out set. Include a paired bootstrap or equivalent comparison between FlowECG at 10–25 NFE and SSSD-ECG at 200 NFE. Also validate the retrained SSSD-ECG baseline against the original published checkpoint/metrics before using it as the baseline. If FlowECG at 10–25 NFE falls within the uncertainty of SSSD-ECG at 200 NFE, the claim stands; otherwise it should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline result is that FlowECG 'maintains generation quality with substantially fewer sampling steps' at 10–25 NFE versus 200 for SSSD-ECG. The only evidence for this is Figure 2, a performance-degradation plot. The text gives numeric degradation for SSSD-ECG (DTW over 600 at 2 NFE and over 200 at 25 NFE) but no numeric values for FlowECG at any reduced NFE, no table, no error bars, no repeated seeds, and no statistical comparison. Table 1 reports only the 200-NFE comparison, and even that table has no variance estimates. If FlowECG's low-NFE curves are within noise of SSSD-ECG's 200-NFE performance, the order-of-magnitude speedup claim is not supported; if the observed effect is driven by a single seed or by an unvalidated retrained baseline, the conclusion could invert. The retrained SSSD-ECG baseline is also not validated against published numbers, so a weak baseline could inflate FlowECG's apparent advantage. These are load-bearing because the central contribution is exactly the low-NFE efficiency result, not the flow-matching framework itself, which is credible and standard.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper adapts the SSSD-ECG architecture to flow matching for conditional 12-lead ECG generation, replacing the diffusion process with a continuous ODE-based flow trained with a standard conditional flow-matching loss. On PTB-XL, Table 1 reports that at 200 neural function evaluations FlowECG improves on MMD, Wasserstein distance, and spectral similarity over a retrained SSSD-ECG baseline, while DTW is slightly worse. The main claim is that FlowECG maintains quality at 10--25 NFE, whereas SSSD-ECG degrades sharply when its 200 evaluation budget is reduced; this is supported only by Figure 2, with no numeric values for FlowECG at reduced NFE.","tokens_in":6916,"tokens_out":3472,"duration_ms":41713,"significance":"If the low-NFE result holds, the contribution is practically valuable: an order-of-magnitude reduction in sampling cost for synthetic ECG generation could enable larger-scale augmentation and real-time deployment. The methodological step is incremental but sensible, reusing a proven architecture with standard flow matching. The paper does not provide code, machine-checked proofs, or parameter-free derivations; the evaluation is empirical, so the strength of the contribution rests entirely on the reliability of the reported comparisons.","major_comments":[{"comment":"The central efficiency claim is not supported by reported numbers. The text gives SSSD-ECG degradation values (DTW > 600 at 2 NFE, > 200 at 25 NFE) but provides no numeric values for FlowECG at any reduced NFE, no table of the NFE sweep, no error bars, no seeds, and no sample counts. A reader cannot tell whether FlowECG's 10--25 NFE curves are within noise of its own 200-NFE performance or of the baseline's 200-NFE performance. This is load-bearing because the headline contribution is exactly 'comparable with 10--25 evaluations.' Please add a table reporting all four metrics for FlowECG and SSSD-ECG at NFE 2, 5, 10, 25, 50, 100, 200, with mean ± std over at least three seeds/evaluation runs and the number of generated and test samples, plus an appropriate statistical comparison for the 10--25 versus 200 NFE claim.","section":"§3.3, Figure 2"},{"comment":"The baseline fairness is not established. The manuscript says the baseline is a 'retrained SSSD-ECG,' but it does not validate the retrained model against published SSSD-ECG results, nor does it report training duration, convergence, or hyperparameter search. If the retrained baseline is weaker than the published model, the 200-NFE comparison and the subsequent NFE comparison could be misleading. Please report the published SSSD-ECG values alongside the retrained ones, or use released checkpoints, and state how the retraining was performed (epochs, early stopping, random seeds, final training loss). This is needed to support the claim that FlowECG outperforms the baseline on three of four metrics.","section":"§3.3, Table 1"},{"comment":"The evaluation protocol is underspecified. The manuscript does not state how many real and generated ECGs were used for the metrics, which PTB-XL split is used, how the conditioning-label distribution is matched, the kernel choice for MMD, or the exact feature set for the Wasserstein metric. These choices can materially change DTW, MMD, and Wasserstein values and affect the comparisons at all NFE levels. Please provide a precise protocol, including sample counts and preprocessing, and consider reporting per-lead results in addition to the averaged values.","section":"§2.3, §3.3"}],"minor_comments":[{"comment":"The heading 'Mean Maximum Discrepancy' should be 'Maximum Mean Discrepancy.'","section":"§2.3"},{"comment":"The flow-matching loss is not fully defined: specify the interpolation x_t = (1-t)x_0 + t x_1 and the target velocity u_t = x_1 - x_0, and clarify the distribution of x_0 and x_1 in the expectation.","section":"Eq. (1)"},{"comment":"The abstract and Section 3.3 are inconsistent about the reduced-NFE claim: the abstract says 10--25 evaluations, while the text says '10-20 NFE.' Please align these statements.","section":"§3.3"},{"comment":"Typo: 'bette capture' should be 'better capture.'","section":"§3.1"},{"comment":"The figure lacks axis labels, a legend, and numeric annotations; given that it is the sole evidence for the central efficiency claim, it should be self-contained and accompanied by the numeric table requested above.","section":"Figure 2"},{"comment":"The abbreviation 'SimScore' appears in Table 1 but is not defined in the text; define it where the spectral similarity score is introduced.","section":"§2.3.3"}],"recommendation":"major_revision","confidential_remarks":"This is a plausible and useful adaptation, but the manuscript is currently not acceptable as a published record of the efficiency claim because the supporting evidence is not quantified. The requested NFE table and baseline validation are within the scope of a revision, so I do not recommend rejection. I would also suggest the editors verify that the venue's expectations for reproducibility (e.g., code/data release) are met, since no code is mentioned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you work on generative models for medical time series and want a quick look at a standard method swap. The new thing: FlowECG replaces the diffusion process in SSSD-ECG with a flow matching objective, keeping the architecture. That combination is new for 12-lead ECG synthesis, and the idea is straightforward—the training loss in Eq. 8 is the standard flow matching loss from Lipman et al., applied to the SSSD net. No invented entities or circular reasoning. That part is solid.\n\nWhat the paper does well: clear description of the method, reasonable choice of metrics (DTW, Wasserstein, MMD, spectral similarity), and an honest comparison against a retrained SSSD-ECG baseline. The 200-NFE numbers, while single-run, are not internally absurd: FlowECG wins on three of four metrics, with DTW worse (73.49 to 83.96). The abstract's \"matches ... outperforming\" is a bit loose; it's mostly within noise likely.\n\nThe soft spots are load-bearing. The headline claim—comparable quality at 10–25 NFE versus 200—rests entirely on Figure 2, which shows degradation curves but no numeric values for FlowECG at any reduced NFE, no error bars, no seeds. The text gives SSSD-ECG numbers at 2 and 25 NFE (DTW over 600 and over 200) but not FlowECG's. If FlowECG's low-NFE curves are within noise of the baseline's 200-NFE performance, the speedup claim is unsupported; if the retrained baseline is weaker than the published one (and it's not validated against any published numbers), the comparison inflates FlowECG's advantage. That matters because the efficiency result is the whole contribution; the flow matching framework itself is credible and standard.\n\nOverall: a plausible engineering paper with a meaningful but unverified central claim. It deserves a reviewer's time, mainly to demand a numeric NFE sweep table, error bars or multiple seeds, baseline validation against published SSSD-ECG numbers, and ideally code. If those come through, the result would be useful for the ECG synthesis community. For now I'd not cite it as evidence for the speedup, but I'd bring it up in a reading group as a case study in how a credible method can be undercut by missing evaluation details.","headline":"Plausible method swap with an under-supported headline: the 10–25 NFE efficiency claim lives in an unquantified figure, and the retrained baseline is not validated.","tokens_in":7429,"tokens_out":1869,"would_cite":false,"duration_ms":20987,"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":"FlowECG shows flow matching can generate 12-lead ECGs with 10–25 neural function evaluations instead of 200, with comparable or better quality.","keywords":["flow matching","ECG synthesis","conditional generation","diffusion models","12-lead ECG","sampling efficiency","medical signal processing"],"falsifier":"A reader could refute the central claim by independently training the same two models on the same public 12-lead ECG data and measuring the four metrics over multiple random seeds and sampling budgets. If the 10–25 step outputs show a clinically significant drop in waveform fidelity, or if the at-200-step advantage over the diffusion baseline reverses under repeated runs, the claim fails. A blinded cardiologist review of real, flow-generated, and diffusion-generated ECGs at matched budgets would settle whether the reduced-step outputs are clinically usable.","tokens_in":6546,"feed_emoji":"🫀","tokens_out":7217,"duration_ms":77705,"temperature":0.7,"pith_summary":"The paper claims that replacing the iterative denoising process in diffusion-based ECG synthesis with continuous flow dynamics produces a generator that matches the diffusion baseline at 200 neural network evaluations and beats it on three of four fidelity metrics. The central advantage is sampling efficiency: 10–25 neural function evaluations suffice for comparable quality, versus 200 for diffusion, a roughly tenfold reduction in compute. If correct, this makes synthetic 12-lead ECG generation practical for real-time clinical settings and large-scale data augmentation without a clinically significant drop in waveform realism.","feed_headline":"FlowECG matches diffusion ECG quality at 10x fewer steps","feed_subtitle":"The flow-based model needs 10–25 neural calls per sample, not 200, so synthetic ECGs become practical in real time.","key_machinery":"The central mechanism is the linear interpolation flow path between Gaussian noise and a real ECG signal, with the network trained to predict the constant velocity pointing from noise to data. Because the path is straight and the target velocity is constant, a simple Euler integration of the learned ODE from t=0 to t=1 produces a sample, and fewer integration steps are needed than diffusion denoising steps because there is no noise schedule to traverse. The architecture is the same structured state-space backbone as the diffusion baseline, so the comparison isolates the training objective and the sampling procedure.","core_discovery":"FlowECG is a flow-matching adaptation of an established diffusion ECG architecture: it keeps the same 36-layer residual network with structured state-space layers and the same diagnostic-label conditioning, but replaces discrete denoising steps with a learned continuous vector field. Training creates linear interpolations between Gaussian noise and real ECG signals and regresses the network to the constant velocity pointing from noise to data; sampling solves the resulting ordinary differential equation from noise to signal. At 200 neural function evaluations, FlowECG improves maximum mean discrepancy, Wasserstein distance, and spectral similarity relative to the retrained diffusion baseline","pith_inferences":["Not tested in the paper: whether the quality margins persist across multiple random seeds and with repeated training runs; the reported numbers are single-run, so the numerical advantage over the baseline may shift under replication.","A testable extension the authors leave implicit is using the deterministic noise-to-signal map learned by flow matching for controlled interpolation between diagnostic conditions, which stochastic diffusion sampling does not offer as cleanly.","The efficiency result plausibly extends to longer or multi-channel physiological recordings beyond the 12-lead ECG, since the same straight-path flow objective has been used successfully for other waveform domains; the paper does not demonstrate this.","A direct clinical falsification route is a blinded cardiologist comparison of real, flow-generated, and diffusion-generated ECGs at matched sampling budgets; the four summary metrics in the paper do not by themselves certify clinical equivalence."],"forward_implications":["ECG synthesis with 10–25 forward passes instead of 200 puts generative ECG models within reach of real-time clinical screening and bedside monitoring, where latency and compute are constrained.","Because each evaluation is one forward pass, the roughly tenfold reduction transfers almost directly to wall-clock and energy savings when producing large synthetic datasets for augmentation or privacy-preserving sharing.","The generated 12-lead ECGs preserve physiologically consistent lead relationships and waveform morphology, so downstream tasks that rely on anatomically coherent multi-lead signals remain supported.","The stability of flow matching under reduced sampling budgets suggests that adaptive sampling strategies—spending more steps on difficult conditioning labels and fewer on easy ones—could work without retraining."],"supporting_citations":[{"why":"Defines the flow matching objective and the ODE sampling procedure that FlowECG adapts for ECG generation.","marker":"[15]"},{"why":"The diffusion model whose architecture and conditioning mechanism FlowECG reuses, and the baseline it must match or beat.","marker":"[7]"},{"why":"Provides the public 12-lead ECG dataset and the train/test split used for training and evaluation.","marker":"[14]"},{"why":"Supplies the four evaluation metrics (dynamic time warping, Wasserstein distance, MMD, spectral similarity) used to compare real and generated ECGs.","marker":"[8]"},{"why":"Gives the spectral similarity formulation used to score frequency-content preservation.","marker":"[16]"},{"why":"Provides the kernel two-sample test underlying the maximum mean discrepancy metric.","marker":"[17]"}],"fun_headline_variants":["FlowECG: diffusion-quality ECG at 1/10 the steps","Flow matching cuts ECG synthesis steps by 10x","FlowECG: diffusion quality at 10x fewer steps","FlowECG: 10x faster ECG synthesis, same quality"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The evaluation assumes that the four summary statistics measured on a single run without error bars adequately capture how good a synthetic ECG is, and that the diffusion model compared against was trained and evaluated fairly; if either assumption is wrong, the speed-up claim still stands but the quality comparison does not.","fun_headline_variants_meta":{"raw":{"variants":["FlowECG: diffusion-quality ECG at 1/10 the steps","Flow matching cuts ECG synthesis steps by 10x","FlowECG: diffusion quality at 10x fewer steps","FlowECG: 10x faster ECG synthesis, same quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000884,"raw_usage":{"total_tokens":3640,"prompt_tokens":717,"completion_tokens":2923,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":2851}},"tokens_in":461,"tokens_out":2923,"duration_ms":25788,"temperature":1.0,"reasoning_tokens":2851,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:13:56.232385+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could refute the central claim by independently training the same two models on the same public 12-lead ECG data and measuring the four metrics over multiple random seeds and sampling budgets. If the 10–25 step outputs show a clinically significant drop in waveform fidelity, or if the at-200-step advantage over the diffusion baseline reverses under repeated runs, the claim fails. A blinded cardiologist review of real, flow-generated, and diffusion-generated ECGs at matched budgets would settle whether the reduced-step outputs are clinically usable.","supporting_citations":[{"cited_title":"Diffusion-based conditional ecg generation with structured state space models","cited_arxiv_id":null,"evidence_quote":"The diffusion model whose architecture and conditioning mechanism FlowECG reuses, and the baseline it must match or beat."},{"cited_title":"Diffecg: A versatile probabilistic diffusion model for ecg signals synthesis","cited_arxiv_id":null,"evidence_quote":"Supplies the four evaluation metrics (dynamic time warping, Wasserstein distance, MMD, spectral similarity) used to compare real and generated ECGs."},{"cited_title":"Spectral similarity measure using frequency spectrum for hyperspectral image classification","cited_arxiv_id":null,"evidence_quote":"Gives the spectral similarity formulation used to score frequency-content preservation."},{"cited_title":"Borgwardt, Malte J","cited_arxiv_id":null,"evidence_quote":"Provides the kernel two-sample test underlying the maximum mean discrepancy metric."}],"review_version":1}