{"id":"bfc58a3f-c51d-43c7-b183-b382cef93b41","arxiv_id":"2603.13293","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"On synthetic cardiac data, FedCVR — a re-implementation of FedAdam with server-side momentum — is reported to reach F1 0.78 / AUC 0.96 under DP (ε≈13.4), beating stateless and other adaptive baselines, though the paper's own numbers conflict internally.","lead":"FedCVR, a federated learning setup for heart-risk prediction, adds privacy noise to hospital updates and uses a server-side momentum optimizer to keep accuracy high, reportedly reaching F1 0.78 and AUC 0.96 under a modest privacy budget. The paper is a readable engineering case study, but its own headline numbers disagree across abstract, tables, and text, and no code or data is released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed statistical superiority rests on N=5 'independent runs' despite §7.2 fixing random_state=42 for all stochastic processes; Table 7's p-values are therefore not meaningful, leaving the central comparative claim unvalidated.","rationale":"The reader's verdict (REJECT) is supported by multiple text-level inconsistencies, and the fixed-seed issue identified here is a concrete, load-bearing flaw directly tied to the paper's central comparative claim. The reader's weakest_assumption focused on the synthetic dataset's fidelity to real clinical data, which is also a valid concern but is somewhat softened by the paper's framing as a proof-of-concept and its explicit limitation statement in §9.4. The seed contradiction is more decisive: it invalidates the statistical machinery used to support the 'statistically outperforms' claim even on the paper's own terms. If the N=5 runs are not independent, the p-values in Table 7 are meaningless, and the conclusion that FedCVR's momentum provides a 'structural prerequisite' for utility recovery lacks quantitative support. This does not require external data or assumptions about real EHRs; it is an internal inconsistency. Therefore I partially agree with the reader's weakest_assumption but identify the fixed-seed statistical flaw as the more load-bearing concern. The verdict remains REJECT (or at least UNVERDICTED), so I recommend UNCHANGED relative to the reader's verdict. No ad hominem is intended; the critique is on the reproducibility logic of the experimental design.","tokens_in":23891,"tokens_out":2485,"duration_ms":28524,"concrete_test":"Rerun the FedCVR-vs-baseline benchmark (Table 7) with five distinct random seeds (e.g., 0, 1, 2, 3, 4) while keeping all other hyperparameters identical, and report the per-seed F1, AUC, and Recall for each method. If the five runs are identical (zero variance), Table 7's ± values and p-values are artifacts of the fixed seed and the statistical-superiority claim is unsupported. If the runs differ, compute the t-test on the new distributions; if the p-values disappear or change materially, the original statistical claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that FedCVR statistically outperforms FedAvg, FedProx, FedAdagrad, and FedYogi under identical privacy constraints—rests on the statistical significance results in Table 7, which report means±std and p-values from 'N=5 independent runs'. However, §7.2 states: 'A random_state=42 was used throughout all stochastic processes, including synthetic data generation, partitioning, and model initialization, ensuring that performance differences are attributable solely to the aggregation strategy and not initialization variance.' If every stochastic process uses the same fixed seed, the five runs are not independent; they are either identical or differ only in ways that are not controlled, and the t-test assumptions are violated. The reported standard deviations and p-values cannot support the claimed statistical superiority. This is not a minor technicality: the paper's headline conclusion—that 'server-side adaptivity is a structural prerequisite'—depends on this comparative evidence. Without valid statistical support, the empirical case reduces to a single-seed comparison on a synthetic dataset whose labels are generated by rule-based logic (§6.3), meaning the AUC≈0.96 and F1 differences could reflect rule-recovery rather than the optimizer's denoising property. The paper itself concedes the synthetic data does not capture real EHR missingness or noise (§9.4), but the seed contradiction is directly verifiable from the text and undermines the quantitative comparison even within the stated experimental setup. The privacy-accounting issue (§5.3.1 uses q=1.0 while §5.4 uses mini-batch subsampling) is a further concern, but the fixed-seed statistical flaw is the most load-bearing because it invalidates the central comparative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FedCVR, a server-side adaptive (Adam-style) aggregation framework for differentially private federated learning (DP-FL) applied to cardiovascular risk prediction. It uses a synthetic 30,000-record dataset distributed across five clients under non-IID partitions, applies local DP-SGD via Opacus, and benchmarks against FedAvg, FedProx, FedCluster, FedAdagrad, and FedYogi. The paper reports an F1 score of 0.78 and AUC 0.96 at a privacy budget of ε≈13.4, claims that FedCVR statistically outperforms the baselines, and concludes that server-side adaptivity is a structural prerequisite for recovering clinical utility under DP noise. The full-text abstract, however, reports the DP F1 as 0.84, which matches the non-private baseline rather than the 0.78 value given in the results.","tokens_in":24165,"tokens_out":10926,"duration_ms":111972,"significance":"If the claims were supported, the paper would provide a useful engineering case study showing that server-side momentum can mitigate DP noise in tabular federated learning. The paper has strengths: it gives a concrete architecture, uses standard libraries (Flower, PyTorch, Opacus), reports hyperparameters in detail, and includes an explicit limitations section. However, the central statistical evidence is invalid because the 'N=5 independent runs' in Table 7 contradict the fixed random_state=42 policy stated in §7.2, and the headline F1 result is internally inconsistent across the abstract and the results sections. The synthetic-data limitation further restricts the scope of any clinical or architectural conclusion. As a result, the paper's significance is not currently established.","major_comments":[{"comment":"Table 7 reports means±std and p-values from 'N=5 independent runs' with a two-tailed t-test, but §7.2 states that 'random_state=42 was used throughout all stochastic processes, including synthetic data generation, partitioning, and model initialization.' If all runs share one seed, they are not independent replicates, so the reported standard deviations and p-values are not meaningful. This invalidates the central claim that FedCVR 'statistically outperforms' the baselines and the contribution titled 'Statistical Validation.' The experiments must be rerun with distinct seeds and the significance analysis redone.","section":"§7.2 and Table 7"},{"comment":"The full-text abstract states that 'integrating server-side momentum as a temporal denoiser allows the architecture to achieve a stable F1-score of 0.84 and an Area Under the Curve (AUC) of 0.96,' but §8.3 reports the F1 stabilizing at 78%, and Table 8 gives F1=0.78±0.03 for σ=1.0 (ε≈13.4), with the non-private baseline at 0.84. The arXiv abstract itself reports 0.78 versus a 0.84 baseline. Thus the full-text abstract reports the baseline value as the DP result. This direct self-contradiction must be corrected and all summaries harmonized.","section":"Abstract vs. §8.3/Table 8"},{"comment":"The 'temporal denoiser' explanation is not established by the experiments. FedYogi also maintains a first moment, so the presence of momentum alone does not distinguish FedCVR from FedYogi. The paper attributes the advantage to the 'specific tuning' of the Adam-based update (§4.3) without an ablation (e.g., varying β1, removing momentum) or a theoretical analysis of the second-moment update rule. Consequently, the conclusion that 'server-side adaptivity is a structural prerequisite' (§10.1) is an overgeneralization from a single synthetic dataset and one privacy configuration.","section":"§4.4, §9.1, Eq. (21), §10.1"},{"comment":"The synthetic dataset's target labels are assigned by rule-based logic derived from clinical guidelines, using the same six features that are given to the model. The paper concedes in §9.4 that the synthetic data 'does not fully capture the missing values and unstructured noise that are typical of raw EHRs.' The headline AUC≈0.96 and the optimizer ranking may therefore reflect recovery of the labeling rule rather than the optimizer's denoising property. The claims of a 'validated engineering blueprint' and clinical utility go beyond the evidence; at minimum the conclusions should be tempered, and ideally the framework should be validated on real EHR data.","section":"§6.3, §9.4, §10.1"},{"comment":"Section 5.3.1 states that Rényi Differential Privacy (RDP) accounting is used to track privacy loss, while §8.6 says 'Future iterations of this work will explore advanced composition techniques, such as Rényi Differential Privacy (RDP).' This contradiction leaves the derivation of ε≈13.4 unclear. The paper should state unambiguously which accounting method was actually used and provide the computation behind the reported privacy budget.","section":"§5.3.1 vs. §8.6"}],"minor_comments":[{"comment":"The noise multipliers are reported as σ∈{0.8,1.1,1.5} in §7.2 but as {0.5,1.0,1.5} in Table 8 and in the abstract. Reconcile these values.","section":"§7.2 vs. Table 8"},{"comment":"Table 2 gives patient counts of 2,487 and 513, totaling 3,000, but the stated total is 30,000. The percentages (82.9%/17.1%) imply counts of 24,870 and 5,130. Correct the arithmetic.","section":"Table 2"},{"comment":"The statement that the model 'converged to a performance plateau nearly identical to the non-private baseline' is inconsistent with Table 8, which shows F1=0.78 vs. 0.84 at σ=1.0. This is a gap of 6 percentage points and should be described accurately.","section":"§9.2"},{"comment":"The statement that 'Data and source code will be made available on request' is insufficient for the reproducibility promised in the contributions, especially given the statistical validation claims. The code and data should be provided with the submission.","section":"Software and Data Availability"},{"comment":"References [13] and [33] appear to be the same paper and are duplicated. Please merge or renumber.","section":"References"},{"comment":"Algorithm 1 uses 'ε' for the numerical stability constant (line 13) and in the Require line, which collides with the privacy budget notation ε used throughout the paper. Rename the stability constant to τ, consistent with Eq. (9).","section":"Algorithm 1"}],"recommendation":"reject","confidential_remarks":"The fixed-seed contradiction is directly verifiable and invalidates the paper's primary significance testing. The abstract inconsistency, the RDP contradiction, and the arithmetic error in Table 2 suggest that the manuscript has not been carefully checked. In my view, the necessary corrections go beyond a normal revision: the experimental campaign would need to be rerun with proper multi-seed replicates and ideally with real-world data. I therefore recommend rejection, though a substantially revised submission with corrected methodology and reporting could be considered on its own merits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a case study of FedAdam (server-side momentum) in differentially private federated learning for cardiovascular risk, evaluated on synthetic data. The idea isn't new—it's FedOpt with stock hyperparameters, and the \"temporal denoiser\" is the standard EMA-averages-out-zero-mean-noise property. What could have been a useful engineering data point is undermined by the paper's own text: the arXiv abstract says DP F1=0.78 vs baseline 0.84, the full-text abstract says DP F1=0.84, §8.3 says non-DP F1 stabilizes at 78%, and Table 8 lists baseline F1 as 0.84. These cannot all be right.\n\nThe more serious problem: Table 7 reports p-values from N=5 independent runs, but §7.2 states random_state=42 was used for all stochastic processes. Runs with the same seed are either identical or not independent, so the t-tests are meaningless. That kills the central claim that FedCVR statistically outperforms FedAvg, FedYogi, and the rest. The privacy accounting also mixes q=1.0 with mini-batch subsampling, and no code or data is provided to check anything.\n\nCredit where it's due: the paper is honest that it's not proposing a new optimizer, it cites FedOpt correctly, and the engineering intuition—momentum as a cheap fix against DP noise—is plausible. The synthetic setup is clean for ablation, and the authors acknowledge in §9.4 that it doesn't capture real EHR noise or missingness. But as submitted, it's a plausible story on rule-based synthetic labels, not a validated blueprint.\n\nWho's this for? Someone wanting a quick example of how to wire FedAdam into Flower/Opacus might skim it, but I wouldn't trust the reported numbers. The paper needs a rewritten experimental protocol with proper seeds, released code/data, and consistent reporting before it's worth referee time. I'd desk reject as is.","headline":"The paper's headline numbers contradict each other and the statistical significance tests are invalid because all runs share one seed; the underlying idea is just FedAdam with a standard EMA argument, so the only real contribution is an untrustworthy benchmark.","tokens_in":24935,"tokens_out":4465,"would_cite":false,"duration_ms":44888,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A differentially private federated learning architecture for cardiovascular risk prediction keeps clinical utility by using server-side momentum to average out privacy noise, reaching AUC 0.96 at a utility-prioritized privacy budget.","keywords":["federated learning","differential privacy","cardiovascular risk prediction","server-side adaptive optimization","momentum","temporal denoising","non-IID data","privacy-utility trade-off"],"falsifier":"Run the same five-way optimizer comparison on a real multi-site EHR cohort under the same DP settings (σ=1.0, 100 rounds, RDP accounting); if FedCVR's AUC gap over FedYogi and FedAvg narrows or reverses, the momentum-denoiser result is an artifact of the synthetic, rule-based benchmark.","tokens_in":23577,"feed_emoji":"🫀","tokens_out":11184,"duration_ms":108142,"temperature":0.7,"pith_summary":"This paper tries to establish that in a federated learning network with local differential privacy, the server-side optimizer — not the client training — is what determines whether the model survives privacy noise. Specifically, it claims that an Adam-style aggregator with momentum acts as a temporal denoiser: averaging each client update across rounds filters out the zero-mean Gaussian noise added for privacy while preserving the true learning signal. On a synthetic cardiovascular risk dataset, FedCVR reaches F1≈0.78 and AUC≈0.96 under a utility-prioritized privacy budget (epsilon≈13.4), statistically beating stateless baselines (FedAvg, FedProx) and other adaptive optimizers (FedAdagrad, FedYogi) under identical privacy constraints. If correct, this reframes the privacy-utility trade-off as an engineering problem: formal privacy guarantees need not force a clinical utility collapse, provided the aggregation layer has memory.","feed_headline":"Restores clinical utility under privacy noise via server momentum","feed_subtitle":"At an epsilon of 13.4, momentum filtering holds AUC at 0.96 and F1 at 0.78, near the non-private baseline.","key_machinery":"The central object is the FedCVR server-side adaptive aggregation step, an Adam-style optimizer with bias-corrected first and second moments. The first moment, m_t = β1 m_{t-1} + (1-β1) g_t, is the temporal filter: since each DP-perturbed update g_t contains true gradient plus zero-mean Gaussian noise, the moving average approximates the expected gradient and damps the high-frequency noise while preserving the low-frequency signal. The second moment v_t and the update w_{t+1} = w_t - η m̂_t/(√v̂_t+τ) scale the step size per parameter. This stateful, low-pass filtering is the mechanism that distinguishes FedCVR from stateless aggregators and from non-momentum adaptive optimizers.","core_discovery":"On its own terms, the paper's central claim is that server-side momentum is the operative mechanism for recovering clinical utility under differential privacy. In the FedCVR update rule, the server maintains first and second moment estimates of the aggregated pseudo-gradient and updates the global model with an adaptive learning rate; the first moment is a moving average that low-pass filters the DP noise superimposed on each client update. The reported validation shows a stable F1-score of about 0.78 and AUC of 0.96 at a cumulative privacy budget of epsilon≈13.4 (delta=1e-5), compared with an F1 of 0.84 for the non-private baseline, and statistically significant superiority over FedAvg, Fed","pith_inferences":["The temporal-denoiser view suggests a testable design rule: any server aggregator that keeps a sufficiently long exponential moving average of updates should recover most of the utility lost to DP noise; future work could systematically vary β1 to map the noise-versus-momentum trade-off.","Because the synthetic labels were generated by a deterministic rule using the same six features the model sees, the AUC 0.96 partly measures rule recovery; on real EHR data with missingness, measurement error, and unmeasured confounders, the optimizer ranking reported here should be re-checked before generalizing to real networks.","Independently recomputing the privacy accountant would strengthen the claim: with full participation (q=1.0) and 100 rounds, standard RDP composition for noise multiplier σ=1.0 may yield an epsilon well above 13.4, in which case the 'operational privacy budget' framing would need revision.","The paper's universal phrasing that server-side adaptivity is a 'structural prerequisite' goes beyond the evidence of a single synthetic tabular setting; it should be read as a mechanism hypothesis, not a proven law."],"forward_implications":["A utility-prioritized privacy budget around epsilon 13.4 can sustain an AUC of 0.96 and F1 ≈0.78, only a small drop from the non-private F1 of 0.84, so formal DP does not automatically destroy clinical viability.","Dropping server-side state (FedAvg, FedProx) costs roughly 0.06–0.08 in AUC under the same DP noise, so aggregation memory, not client regularization, is the key to robustness.","Momentum-based aggregation outperforms other adaptive optimizers (FedAdagrad, FedYogi) under identical privacy constraints, indicating that the specific first-moment update—not adaptivity in general—is the noise-filtering component.","Under a stricter budget (epsilon≈6.6), utility degrades gracefully (F1≈0.65) rather than collapsing, supporting deployment in trusted consortium settings.","Communication cost is unchanged relative to FedAvg; only server memory grows (storing w, m, v), making the mechanism cheap to add to existing federated learning systems."],"fun_headline_variants":["Server momentum restores accuracy in private heart-risk prediction","Server momentum protects heart-risk models from privacy noise","Private federated learning keeps heart-risk AUC at 0.96 with momentum","Momentum recovers clinical utility in differentially private federated learning","Server-side momentum beats stateless baselines for private heart-risk AI"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The synthetic dataset, whose high-risk labels were assigned by a deterministic rule from the same six features the model is given, faithfully represents real clinical records; if real EHR noise and missingness change the ranking, the architectural conclusions can collapse.","fun_headline_variants_meta":{"raw":{"variants":["Server momentum restores accuracy in private heart-risk prediction","Server momentum protects heart-risk models from privacy noise","Private federated learning keeps heart-risk AUC at 0.96 with momentum","Momentum recovers clinical utility in differentially private federated learning","Server-side momentum beats stateless baselines for private heart-risk AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2506,"prompt_tokens":817,"completion_tokens":1689,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1603}},"tokens_in":561,"tokens_out":1689,"duration_ms":13335,"temperature":1.0,"reasoning_tokens":1603,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:52:45.694208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same five-way optimizer comparison on a real multi-site EHR cohort under the same DP settings (σ=1.0, 100 rounds, RDP accounting); if FedCVR's AUC gap over FedYogi and FedAvg narrows or reverses, the momentum-denoiser result is an artifact of the synthetic, rule-based benchmark.","supporting_citations":[],"review_version":1}