{"id":"c45825b6-885e-4635-981b-04ba53e79e0b","arxiv_id":"2607.08595","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FedAvg DeepSurv across Lifelines (n=148k, self-report) and Rotterdam Study (n=10k, linked outcomes) raised C-statistics from 0.728 to 0.739 and 0.783 to 0.787 versus local training.","lead":"Federated DeepSurv training across Lifelines and the Rotterdam Study raised CVD risk C-statistics without sharing patient records. The work shows a large self-report cohort can still help a smaller clinical-outcome cohort under privacy constraints.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Overlapping CIs and size-weighted FedAvg leave the Rotterdam gain statistically and causally under-supported.","rationale":"The reader correctly flags the informativeness of Lifelines labels under sample-size-weighted FedAvg as the weakest assumption. That concern is load-bearing: the only quantitative support for the strongest claim is a 0.011 point-estimate lift whose CIs completely overlap and for which no test of the difference is given, while the aggregation rule and convergence trajectory both favor the larger, noisier cohort. The paper is otherwise transparent and the methods are standard, so the contribution remains accept-shaped if the authors supply the missing formal comparison (and ideally a quality-aware or early-stopped ablation). No stronger internal inconsistency or fabrication issue appears; the soft spot is statistical and causal under-support of the reported gain on the primary cohort. Verdict therefore stays CONDITIONAL, aligned with the reader.","tokens_in":12141,"tokens_out":588,"duration_ms":6071,"concrete_test":"On the same 10 stratified splits, compute the per-run C-statistic difference (federated − local) for Rotterdam and apply the corrected resampled t-test (Nadeau & Bengio) already used for the CIs; report the mean delta, its 95% CI, and p-value. If that CI includes zero (or p>0.05), the headline improvement is not statistically supported. Optionally re-run FedAvg with equal node weights or early-stop at the Rotterdam peak (~iter 12) and re-evaluate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that FedAvg of DeepSurv across Lifelines and Rotterdam raises Rotterdam C-statistic from 0.728 (0.717–0.739) to 0.739 (0.728–0.749). Table 2 and the Results text show the federated point estimate equals the upper bound of the local CI and the local point equals the lower bound of the federated CI; the intervals fully overlap. No paired or corrected-resampled test of the difference is reported, only separate CIs. Because FedAvg (Eq. 4) weights by sample size, Lifelines (~15× larger) dominates the aggregate; the Discussion itself notes a late-iteration decline on Rotterdam after an early peak near iteration 12 (Figure 2A). Thus the modest gain could be noise, transient, or an artifact of size-weighted averaging of a noisier, younger, lower-event-rate cohort (self-report + midpoint times, Section 2.2.2) rather than genuine transfer of useful signal. Without a formal test of the delta and without an ablation that equalizes or quality-weights the two nodes, the claim that federation improves the primary evaluation cohort remains only weakly secured.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript evaluates federated training of DeepSurv (FedAvg over 20 rounds on Vantage6) across two heterogeneous Dutch cohorts—Lifelines (N=148,230; self-reported CVD with midpoint event times) and the Rotterdam Study (N=10,155; digitally linked clinical outcomes)—without sharing individual-level data. After FHIR-based harmonization of 12 standard CVD predictors, local vs federated discrimination is compared on held-out stratified splits (10 runs, corrected resampled t-test CIs). The primary claim is that federation raises the C-statistic relative to local training: Rotterdam 0.728 (0.717–0.739) → 0.739 (0.728–0.749); Lifelines 0.783 (0.775–0.791) → 0.787 (0.780–0.792), with Rotterdam as the primary evaluation cohort because of complete follow-up.","tokens_in":12427,"tokens_out":1396,"duration_ms":23957,"significance":"If the performance gain is real and transferable, the work is a useful real-world demonstration that federated deep survival models can improve CVD risk prediction across cohorts that differ in size, age, event rate, and outcome ascertainment, while remaining within existing GDPR/governance constraints. Strengths include a genuine multi-site (not pseudo-split) setup, a documented HL7 FHIR harmonization pipeline, fixed architecture/seed, 10 stratified runs with corrected resampled CIs, primary evaluation on the better-labeled cohort, and an unusually candid discussion of size imbalance and label noise. The contribution is incremental relative to the prior MyDigiTwin infrastructure paper, but the multi-cohort evaluation is the right next step for the field.","major_comments":[{"comment":"Table 2 and §3: The central claim that federation improves discrimination rests on point-estimate increases whose 95% CIs fully overlap (Rotterdam federated point equals the local upper bound; local point equals the federated lower bound). Only separate CIs for each setting are reported; no paired or corrected-resampled test of the federation-minus-local difference is given. Without a formal test of ΔC (or equivalent), the primary claim remains statistically under-supported and should be either tested or substantially softened in the abstract and conclusions.","section":"Table 2 / Results §3"},{"comment":"Eq. (4) and Figure 2(A): FedAvg weights updates by sample size, so Lifelines (~15× larger training set) dominates the aggregate. The Rotterdam trajectory peaks near iteration 12 (~0.740) then declines slightly by iteration 20, which the Discussion attributes to growing Lifelines influence. Reporting only the final (iteration-20) model as the federated result, without a pre-specified stopping rule, quality-weighted aggregation, or an ablation that equalizes or down-weights the noisier node, leaves open whether the reported Rotterdam gain is stable transfer of signal or a transient/size-weighting artifact. At minimum, report best-iteration and last-iteration results with a formal comparison, and discuss (or ablate) alternatives to pure nk/n weighting given the known label-quality asymmetry (Section 2.2.2).","section":"Eq. (4) / Figure 2(A) / Discussion"},{"comment":"Section 2.2.2 and Discussion: The weakest modeling assumption is that Lifelines self-reported events with midpoint-interval times remain informative under size-weighted FedAvg for a model evaluated on Rotterdam’s precise clinical outcomes, despite large differences in age, event rate (2.6% vs 13.2%), and outcome definition. The paper acknowledges this but does not quantify sensitivity (e.g., event-time noise simulation on Lifelines, or federation with event-type subsets). A sensitivity analysis or explicit bound on how much label noise can be tolerated would make the transfer claim more credible.","section":"Section 2.2.2 / Discussion"}],"minor_comments":[{"comment":"Abstract and Table 2: State explicitly that CIs overlap and that no formal test of the difference was performed (or add the test), so readers are not left to infer statistical significance from non-overlapping language that is not used but easily misread.","section":"Abstract / Table 2"},{"comment":"Section 2.3.1: The architecture [12, 16, 16, 1], dropout 0.20, lr 1e-4, batch 1024, 100 local epochs, and 20 rounds are free hyperparameters with no tuning description. Briefly note whether they were fixed a priori from the prior MyDigiTwin work or selected on validation data.","section":"Section 2.3.1"},{"comment":"Section 2.3.3: Median imputation is done independently per site; a short note on missingness rates by cohort would help assess whether imputation differences could affect federation.","section":"Section 2.3.3"},{"comment":"Discussion: Competing risks (non-CVD death treated as censoring) is correctly flagged as a limitation for absolute risk in older Rotterdam participants; consider stating that discrimination (C-statistic) is the sole reported metric and is less directly affected than calibration/absolute risk.","section":"Discussion"},{"comment":"References / Katzman et al.: The DeepSurv citation line appears garbled in the reference list (“Ded2008generalepsurv”); please correct the bibliographic entry.","section":"References"},{"comment":"Figure 2: Add error bars or shaded bands for the 10-run variability at each iteration so the trajectory can be judged against run-to-run noise.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"Fit and novelty are appropriate for a methods/application venue in healthcare ML, but the advance over the authors’ own MyDigiTwin infrastructure paper is mainly the two-cohort evaluation. The statistical support for the headline gain is the main editorial risk: if the authors cannot produce a significant ΔC or a convincing sensitivity analysis, the abstract claims should be dialed back to “feasibility with modest, non-significant point-estimate gains.” I would not reject on that basis alone if revised carefully."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new piece here is not the algorithm—DeepSurv + FedAvg + Vantage6 + FHIR is already in their MyDigiTwin infrastructure paper—but a genuine two-site run on Lifelines (N≈148k, self-report, midpoint event times) and Rotterdam (N≈10k, digitally linked outcomes). That heterogeneity is the actual contribution.\n\nThey do the basics cleanly. Ten stratified resamples, corrected resampled t-test CIs, fixed architecture and seed, primary evaluation on the better-labeled cohort, and transparent trajectories in Figure 2. Point estimates move the right way on both sites (Rotterdam 0.728→0.739; Lifelines 0.783→0.787). The Discussion is unusually candid about size imbalance, late-iteration dip on Rotterdam after an early peak near round 12, and the noisiness of self-report labels. That honesty is a plus.\n\nThe soft spots are real but proportionate. The CIs fully overlap; the federated Rotterdam point sits exactly on the upper edge of the local CI and vice versa. They never report a formal test of the delta. Because FedAvg weights by sample size, Lifelines dominates, so the modest Rotterdam lift could be noise, transient, or partial transfer of a younger, lower-event-rate signal rather than clean knowledge transfer. Competing risks are ignored (non-CVD death treated as censoring in an older cohort). Only two nodes. No public code or data, which is expected for these cohorts but still limits re-runnability. None of these kill the paper; they just keep the claim modest.\n\nThis is for people who actually build multi-site medical FL or CVD risk models under GDPR constraints. It is not a methods breakthrough and will not change clinical practice. It is a careful empirical data point that federation can still help when one cohort is large-and-noisy and the other is small-and-clean.\n\nI would send it to peer review. A serious referee can demand the paired test, a quality- or equal-weight ablation, and a competing-risk note; the core experiment is solid enough to deserve that conversation. Worth reading if you work in this niche; not required reading otherwise.","headline":"Real two-cohort federated DeepSurv result with honest modest gains; the Rotterdam improvement is directionally consistent but statistically soft under size-weighted FedAvg and overlapping CIs.","tokens_in":13098,"tokens_out":542,"would_cite":false,"duration_ms":5387,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Federated deep survival models trained across two unmatched Dutch cohorts beat local-only training on cardiovascular risk without sharing patient records.","keywords":["Federated learning","Cardiovascular disease","Healthcare AI","Deep survival neural networks","Lifelines","Rotterdam Study","DeepSurv","FedAvg"],"falsifier":"Re-run the same 10-split protocol after deliberately dropping or randomly flipping a substantial fraction of Lifelines event labels and check whether the Rotterdam C-statistic still rises above its local baseline of 0.728; if the gain disappears, the claim that the large noisy cohort helps the precise one fails.","tokens_in":13027,"feed_emoji":"❤️","tokens_out":853,"duration_ms":8179,"temperature":0.7,"pith_summary":"Cardiovascular risk tools are usually built on one site's data or on pooled individual records, both of which run into privacy barriers when institutions want to collaborate. This paper shows that a deep survival network can be trained jointly on two real population cohorts that never exchange raw data: the large Lifelines study with self-reported events and the smaller Rotterdam Study with digitally linked clinical outcomes. After twenty rounds of federated averaging, discrimination rose in both cohorts, with the clearer gain on the Rotterdam side that has complete follow-up. The result matters because it suggests a practical route to larger, more generalizable risk models while leaving patient-level data behind institutional walls.","feed_headline":"Federated heart-risk models beat local training without sharing data","feed_subtitle":"Two Dutch cohorts improve each other's C-statistics while patient records stay on site","key_machinery":"Horizontal Federated Averaging (FedAvg) of DeepSurv weights: each site trains the same [12,16,16,1] network for 100 local epochs, then a central aggregator computes a sample-size-weighted average of the weights and redistributes it for the next of 20 rounds; raw records never leave their sites.","core_discovery":"DeepSurv models trained with Federated Averaging across Lifelines (n=148,230, self-reported outcomes) and the Rotterdam Study (n=10,155, digitally linked outcomes) achieve higher C-statistics than identical models trained only on local data: Rotterdam rises from 0.728 to 0.739 and Lifelines from 0.783 to 0.787, with primary evaluation on Rotterdam because of its complete follow-up.","pith_inferences":["Adaptive early-stopping or quality-aware reweighting of client updates would likely reduce the late-training dip seen on the smaller, higher-quality node.","The same harmonization-plus-FedAvg pipeline should transfer to other time-to-event endpoints (cancer, dementia) that face identical privacy and label-noise tensions.","If competing-risk models replace simple censoring of non-CVD death, absolute risk calibration in older Rotterdam participants would be the natural next stress test."],"forward_implications":["Institutions that cannot pool patient records can still improve local CVD risk models by exchanging only model weights under FedAvg.","A large cohort with imperfect outcome ascertainment can still raise discrimination on a smaller cohort that has gold-standard follow-up.","Convergence must be monitored site-by-site; sample-size-weighted averaging can produce non-monotonic trajectories when cohort sizes differ by an order of magnitude.","Existing GDPR-compliant cloud workspaces and joint-controller agreements already support this style of multi-cohort deep survival training."],"fun_headline_variants":["Federated DeepSurv lifts CVD C-stats without sharing patient records","Privacy-preserving federation beats local heart-risk models on two cohorts","Cross-cohort federated learning raises Rotterdam CVD C-stat to 0.739","Federated Averaging improves CVD risk prediction over solo institutional training","Heterogeneous Dutch cohorts boost each other's heart-risk C-stats via federation"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That Lifelines' noisier, self-reported events with midpoint-interval dates still carry enough signal under sample-size-weighted averaging to improve a model judged on Rotterdam's precise clinical outcomes, despite large gaps in age, event rate, and outcome definition.","fun_headline_variants_meta":{"raw":{"variants":["Federated DeepSurv lifts CVD C-stats without sharing patient records","Privacy-preserving federation beats local heart-risk models on two cohorts","Cross-cohort federated learning raises Rotterdam CVD C-stat to 0.739","Federated Averaging improves CVD risk prediction over solo institutional training","Heterogeneous Dutch cohorts boost each other's heart-risk C-stats via federation"]},"model":"grok-4.5","effort":"low","cost_usd":0.005856,"raw_usage":{"total_tokens":1536,"prompt_tokens":840,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":58560000,"prompt_tokens_details":{"text_tokens":840,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":597,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":840,"tokens_out":99,"duration_ms":5452,"temperature":1.0,"reasoning_tokens":597,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T04:36:41.348245+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same 10-split protocol after deliberately dropping or randomly flipping a substantial fraction of Lifelines event labels and check whether the Rotterdam C-statistic still rises above its local baseline of 0.728; if the gain disappears, the claim that the large noisy cohort helps the precise one fails.","supporting_citations":[],"review_version":1}