{"id":"6e69d263-73fd-460f-ab26-27421df26d84","arxiv_id":"2506.15626","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Federated learning trained brain-age models on stroke MRI data across 16 centers without centralizing images, beating single-site models and yielding BrainAGE scores associated with diabetes and three-month functional outcome.","lead":"This study tested whether brain-age models can be trained across 16 hospitals without sharing patient MRI data, using federated learning. It found federated training beat single-hospital models and produced brain-age scores that predicted recovery after stroke, though slightly less accurately than pooling all data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Clinical claims rest on unmasked FLAIR BrainAGE; acute lesion burden may drive the reported associations, making the clinical informativeness of FL-derived BrainAGE unestablished.","rationale":"I read the paper's central claim as having two parts: (a) FL can train BrainAGE models on decentralized stroke data with performance close to centralized learning, and (b) the resulting BrainAGE is clinically informative. Part (a) is supported by the reported prediction-error experiments; the comparison to single-site training is weakened by confounding with data volume and cross-center domain shift, but this does not threaten the FL-versus-centralized comparison, which is the core algorithmic claim. Part (b) is the load-bearing condition for the paper's clinical relevance. The FLAIR images are not lesion-masked and the onset-to-imaging delay is unknown, so acute lesion effects are a plausible alternative explanation for the BrainAGE associations with diabetes and functional outcome. The authors acknowledge this in Section 4.5 but provide no quantitative check. Because the clinical value of the biomarker is a major advertised contribution, this is the single most load-bearing concern. The proposed lesion-volume adjustment is feasible with the existing imaging data and would settle whether the associations are confounded. I agree with the reader's weakest_assumption and do not recommend a change to the conditional verdict pending this sensitivity analysis.","tokens_in":17454,"tokens_out":12891,"duration_ms":152148,"concrete_test":"Segment acute/subacute lesions on the 1023 test FLAIR images with an automated stroke lesion segmentation tool; (1) test whether lesion volume differs between DM and non-DM patients and between good and poor outcome groups; (2) re-fit the Table 3 logistic regression models adding lesion volume (and lesion location if feasible) as a covariate; (3) optionally recompute BrainAGE after masking FLAIR lesions and repeat the phenotype and outcome analyses. If the DM difference and BrainAGE odds ratios attenuate to non-significance, the clinical claims are confounded by lesion burden; if they persist, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central clinical conclusion—that FL-derived BrainAGE is as informative as centralized BrainAGE—depends on BrainAGE reflecting pre-existing brain aging rather than acute stroke lesion effects. BrainAGE is computed from FLAIR images that are not lesion-masked (Sections 2.3 and 2.4.2.1), and the time from symptom onset to MRI is unknown (Section 4.5). In acute stroke, FLAIR hyperintensity and edema can increase apparent brain age. If lesion volume correlates with diabetes and with 3-month functional outcome, the reported between-group BrainAGE differences (Section 3.4) and odds ratios (Table 3) could be confounded. The logistic regressions adjust for NIHSS, P2P, IVT, and RECA, but NIHSS is a clinical severity score and not a substitute for lesion volume or location. Because both centralized and FL models are trained on the same unmasked data, a lesion confound would not affect the algorithm comparison, but it would invalidate the claim that these BrainAGE values are clinically informative biomarkers of brain health. The paper acknowledges this risk in Section 4.5 but provides no sensitivity analysis, so the clinical claim is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains FLAIR-based BrainAGE models under three data-management strategies—centralized learning, federated learning with FedAvg, and single-site learning—on 1674 ischemic stroke patients treated with mechanical thrombectomy across 16 centers. The test set is fixed to 1023 images from Centers 2–16; centralized and federated models are evaluated with five-fold cross-validation, while single-site models are trained only on Center 1. The authors compare mean absolute age-prediction errors with paired Wilcoxon tests, test BrainAGE differences across vascular risk factors and functional outcome groups with Mann–Whitney tests, and fit logistic regressions for good versus poor three-month functional outcome adjusted for demographic, vascular, and treatment variables. The main reported findings are that centralized models have the lowest prediction errors, federated models are close to centralized and significantly better than single-site models, diabetes and poor functional outcome are associated with higher BrainAGE across models, and BrainAGE has significant odds ratios below one in the outcome regressions. The paper concludes that FL is a viable privacy-preserving approach to BrainAGE and that FL-derived BrainAGE may be as clinically informative as centralized BrainAGE.","tokens_in":17697,"tokens_out":7963,"duration_ms":94040,"significance":"If the conclusions hold, the paper is a useful empirical demonstration of FL for BrainAGE on real multicenter clinical stroke data rather than on simulated partitions. The study has notable strengths: a genuinely multicenter cohort, a fixed test set outside the largest center, five-fold cross-validation for the centralized and federated arms, nonparametric paired tests, several model complexities, and public code and radiomics parameters. The main quantitative result—that FL substantially reduces error relative to a single-site model trained only on Center 1—is plausible and supported by Figure 3. However, the clinical-informativeness claim rests on an unaddressed lesion confound, and the FL-versus-single-site comparison is weakened by the fact that the single-site model is evaluated out-of-distribution and with less data. The secondary clinical analyses also lack multiple-comparison correction. These issues do not invalidate the algorithmic comparison, but they do limit the strength of the conclusions as currently written.","major_comments":[{"comment":"The clinical conclusion depends on BrainAGE reflecting pre-existing brain health, but the FLAIR images are not lesion-masked and the time from symptom onset to imaging is often unknown. Acute FLAIR hyperintensity and edema can inflate predicted age, and lesion burden is plausibly correlated with diabetes and with three-month functional outcome; adjusting for NIHSS does not capture lesion volume or location. The paper acknowledges this in Section 4.5 but provides no sensitivity analysis, such as adding lesion volume as a covariate, excluding patients with large lesions, or recomputing BrainAGE on lesion-masked images. Without such an analysis, the conclusion that FL-derived BrainAGE is 'as clinically informative' as centralized BrainAGE is unsupported. I request either a sensitivity analysis along these lines or a substantially weakened clinical claim that restricts the conclusions to associations observed in this cohort.","section":"§3.4, §3.5, §4.5"},{"comment":"The 'single-site' baseline trains only on Center 1 and is evaluated exclusively on images from Centers 2–16, so the FL-versus-single-site comparison conflates larger training data with distribution shift across scanners and sites. The paper acknowledges this in Section 4.5, but the conclusion still emphasizes 'FL consistently outperformed single-site training.' A held-out evaluation within Center 1, or a single-site baseline trained on a matched amount of data, would better isolate the effect of multicenter data. As it stands, the comparison is best described as an ablation demonstrating the value of multicenter data rather than as a general superiority claim over locally trained models.","section":"§2.3.1, §3.3"},{"comment":"The comparison between centralized and federated models is confounded by training-configuration choices. For the volume and radiomics linear models, centralized and single-site models use inverse-scaling learning-rate decay with L2 weights tuned by cross-validation, while federated models use linear learning-rate decay and adopt L2 weights from single-site training. For VoxelsCNN, centralized and single-site models use Adam for 1000 epochs with BatchNorm, while the federated model uses SGD for 500 epochs with LayerNorm. These choices are motivated by empirical stability, but the observed MAE differences between centralized and federated models could partly reflect hyperparameters or architecture rather than the data-management strategy itself. Please provide a matched-setting ablation, or explicitly discuss the extent to which these differences can affect the rank-ordering in Figure 3.","section":"§2.3.2, §2.3.3"},{"comment":"The clinical association arm performs many univariate tests—six phenotypes across four models and three training configurations, plus twelve good-versus-poor outcome comparisons—and twelve logistic regressions, all without multiple-comparison correction. Several reported significant effects, such as hypertension for some models and sex differences whose direction depends on the model, could be false positives under correction. I recommend reporting FDR-corrected p-values or pre-specifying a small number of primary analyses. The headline findings for diabetes and functional outcome have very small p-values and would likely survive correction, but the secondary phenotype results should be described as exploratory.","section":"§3.4, §3.5, Table 3, Table B.1"}],"minor_comments":[{"comment":"The methods section describes logistic regression with adjusted odds ratios, but Table B.1 is labeled 'Standardized odds ratio'; please clarify whether predictors were standardized and define the standardization, because Table 3 reports unstandardized odds ratios for BrainAGE.","section":"§2.4.2.3, Table B.1"},{"comment":"The sentence 'it is appropriate to compare federated models with centralized models trained on smaller datasets' appears to refer to the single-site models rather than to centralized models; please rephrase to avoid ambiguity.","section":"§4.5"},{"comment":"There is inconsistent spelling between 'centralized' in the text and 'centralised' in the figure captions; please unify the spelling.","section":"Figure 3 and Figure 2"},{"comment":"For the Wilcoxon signed-rank tests, reporting effect sizes or median paired differences would help quantify the magnitude of the differences beyond p-values.","section":"§3.3"},{"comment":"The variable P2P is defined as the interval between MRI and arterial puncture, but Section 4.5 refers to 'the time from imaging completion to treatment initiation'; please ensure the terminology is consistent.","section":"§2.2 and §4.5"}],"recommendation":"major_revision","confidential_remarks":"The main risk to the manuscript is the lesion confound in the clinical arm. If the authors can supply a sensitivity analysis—or explicitly restrict their clinical conclusions to associations without a biological-brain-health interpretation—the paper would be suitable. The single-site comparison issue is acknowledged by the authors and is partly a matter of framing, but should be addressed in the revision. The code release and real multicenter setting are assets that make the paper a reasonable fit for a medical-imaging or applied-ML journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the short version. This is the first genuinely multicenter FL evaluation of BrainAGE on FLAIR images from stroke patients, 16 centers, 1674 patients, real site-specific variability rather than simulated partitions. The ML comparison is careful: fixed test set outside Center 1, five-fold CV for centralized and federated, same models, public code. The main result—FL consistently beats single-site training, though centralized is best—holds up as an empirical claim.\n\nWhat's new: prior FL brain-age work used healthy cohorts or simulated splits. This applies FedAvg to a real clinical cohort and tests whether FL-derived BrainAGE retains associations with diabetes and functional outcome. The association results are consistent across all four models, which is reassuring. The logistic regressions adjust for the obvious confounders.\n\nSoft spots, in order of seriousness. First, the clinical biomarker claim is undermined by the lack of lesion masking. BrainAGE is computed from unmasked FLAIR, and time from onset to imaging is unknown. Lesion volume and edema could inflate predicted age, and lesion burden plausibly correlates with diabetes and with 3-month outcome. The authors acknowledge this in Section 4.5 but run no sensitivity analysis, so the claim that these BrainAGE values are clinically informative of brain health is unsupported. This does not damage the algorithm comparison, since all training configurations share the same images, but it limits what you can say about the biomarker. Second, no multiple-comparison correction across the phenotype and outcome tests. A dozen or so tests per model; some nominally significant results are likely chance. Minor in the sense that the diabetes and outcome associations are robust across all models, but the model-specific HTN/sex findings should be read as exploratory. Third, the single-site comparison is largely a data-volume comparison. The authors address this in 4.5 and cite Stripelis; I agree with their justification—it's the right baseline for the question, but don't oversell it.\n\nOverall: solid empirical application, clean enough ML evaluation, honest limitations. The clinical conclusion is a step ahead of the evidence. Worth a serious referee, and I'd cite it when writing about FL in clinical imaging, with the lesion caveat noted. Bring it to the reading group if you care about FL or BrainAGE.","headline":"A real first: FedAvg on a 16-center clinical stroke cohort for FLAIR BrainAGE, with the expected caveat that the clinical associations are not yet clean of lesion confounds.","tokens_in":18234,"tokens_out":1757,"would_cite":true,"duration_ms":20366,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Federated learning can train FLAIR-based BrainAGE models on real multicenter stroke data without centralizing images, with accuracy and clinical associations close to centralized training and better than single-site training.","keywords":["federated learning","brain age gap","BrainAGE","FLAIR MRI","ischemic stroke","functional outcome","mechanical thrombectomy","multicenter study"],"falsifier":"Compare federated BrainAGE associations before and after masking stroke lesions on the FLAIR images, or in a subgroup with recorded symptom-onset-to-MRI times; if the diabetes and functional-outcome associations disappear after masking or track time-since-onset, the biomarker is capturing acute lesion burden rather than brain age.","tokens_in":17276,"feed_emoji":"🧠","tokens_out":6460,"duration_ms":62445,"temperature":0.7,"pith_summary":"The paper tests whether federated learning can train brain-predicted age difference (BrainAGE) models from FLAIR MRI scans of acute ischemic stroke patients when patient data cannot leave their hospitals. Across 1,674 patients from 16 stroke centers, models trained with the FedAvg algorithm consistently beat models trained on a single site, and approached the accuracy of centralized training on pooled data. The authors also show that BrainAGE derived under federated training keeps its clinical signal: it is higher in patients with diabetes and in patients with poor three-month functional outcomes, and it remains a significant predictor after adjusting for age, stroke severity, and treatment variables. If the results hold, federated learning offers a privacy-preserving route to build and deploy BrainAGE-based prognostic tools in routine stroke care without constructing a central imaging repository.","feed_headline":"Federated learning beats single-site MRI brain-age models","feed_subtitle":"Across 16 centers, federated BrainAGE flagged diabetes and poor recovery without pooling scans.","key_machinery":"The machinery is BrainAGE, the difference between a brain's predicted age and its chronological age, computed from FLAIR MR images after a bias-adjustment step that regresses the predicted-age difference on chronological age. The training side rests on the FedAvg federated algorithm: each hospital center trains locally on its own images for one epoch, sends weight updates to a server, and the server averages them, so raw images never leave the site. Four model families, ranging from a simple volume-based linear regressor to a 3D convolutional neural network, test the approach across computational budgets, and the comparison to single-site training isolates the value of multicenter data.","core_discovery":"The central discovery is that federated learning can produce FLAIR-based BrainAGE estimates on real multicenter stroke data whose clinical associations match those of centralized training. In head-to-head comparisons, centralized learning achieved the lowest absolute age-prediction errors, but federated models were statistically indistinguishable from centralized models for the two simpler volume-based regressors and consistently outperformed single-site models across all four architectures. BrainAGE computed from the federated models was significantly higher in patients with diabetes and in patients with poor functional outcomes at three months, and logistic regressions adjusting for age, sex, vascular risk factors, stroke severity, imaging-to-puncture time, thrombolysis, and recanalization gave BrainAGE odds ratios below 1 in every configuration. The authors conclude that FL-derived BrainAGE may be as clinically informative as centralized BrainAGE.","pith_inferences":["A testable next step would lesion-mask the FLAIR images or record symptom-onset-to-imaging time; if BrainAGE's association with functional outcome weakens, part of the signal is acute lesion burden rather than pre-existing brain aging.","The comparison suggests that FL's benefit grows with model complexity and input dimensionality, since radiomics and CNN models showed the largest single-site degradation, making high-dimensional biomarkers the most promising federated targets.","If confirmed on cohorts with more heterogeneous age distributions across centers, FL-based BrainAGE could double as a harmonization tool, reducing the need for explicit cross-site image normalization.","The odds-ratio pattern hints that age-prediction accuracy and clinical relevance are related but not identical, so model selection for prognosis should use clinical association rather than age error alone."],"forward_implications":["Federated learning lets stroke networks train BrainAGE models on data from many hospitals without a data-sharing agreement that pools images.","FLAIR-based BrainAGE trained under FL remains a significant predictor of three-month functional outcome after adjustment for age, stroke severity, and treatment variables.","Diabetes is the vascular risk factor most robustly linked to higher BrainAGE, since the association appears in all models and both training strategies.","For small or compute-limited centers, FL gives access to multicenter training; the simpler volume-based federated models reach centralized-level accuracy."],"supporting_citations":[{"why":"Supplies the FLAIR radiomics BrainAGE approach and the functional-outcome association that this study extends to federated settings.","marker":"[7]"},{"why":"Provides the FedAvg algorithm used for federated training across hospital centers.","marker":"[20]"},{"why":"Defines the bias-adjustment procedure used to compute BrainAGE from predicted-age differences.","marker":"[32]"},{"why":"Supplies the segmentation method that yields the volume features and white-matter masks used by the models.","marker":"[21]"},{"why":"Provides the 3D convolutional neural network architecture adapted for voxel-based age prediction.","marker":"[27]"},{"why":"Prior federated neuroimaging study used to justify comparing FL against centralized and single-site training.","marker":"[16]"}],"fun_headline_variants":["Federated brain-age rivals centralized, beats single-site","Privacy-preserving brain-age matches pooled stroke MRI","Federated learning: brain-age predicts recovery without pooling","Diabetes and poor outcomes flagged by federated brain-age"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The clinical conclusions stand on the assumption that BrainAGE measures pre-existing brain aging rather than acute stroke lesion effects, since the FLAIR images were not lesion-masked and the time from symptom onset to imaging was often unknown.","fun_headline_variants_meta":{"raw":{"variants":["Federated brain-age rivals centralized, beats single-site","Privacy-preserving brain-age matches pooled stroke MRI","Federated learning: brain-age predicts recovery without pooling","Diabetes and poor outcomes flagged by federated brain-age"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000379,"raw_usage":{"total_tokens":2048,"prompt_tokens":1016,"completion_tokens":1032,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":968}},"tokens_in":632,"tokens_out":1032,"duration_ms":11789,"temperature":1.0,"reasoning_tokens":968,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:52:31.813256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare federated BrainAGE associations before and after masking stroke lesions on the FLAIR images, or in a subgroup with recorded symptom-onset-to-MRI times; if the diabetes and functional-outcome associations disappear after masking or track time-since-onset, the biomarker is capturing acute lesion burden rather than brain age.","supporting_citations":[{"cited_title":"Bretzner, A","cited_arxiv_id":null,"evidence_quote":"Supplies the FLAIR radiomics BrainAGE approach and the functional-outcome association that this study extends to federated settings."},{"cited_title":"McMahan, E","cited_arxiv_id":null,"evidence_quote":"Provides the FedAvg algorithm used for federated training across hospital centers."},{"cited_title":"Beheshti, S","cited_arxiv_id":null,"evidence_quote":"Defines the bias-adjustment procedure used to compute BrainAGE from predicted-age differences."}],"review_version":1}