{"id":"80384536-ddb5-4c8f-a494-2e4edfe192e3","arxiv_id":"2501.10153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A region-wise stacking ensemble improves brain age prediction from structural MRI compared to regional mean gray matter volume.","lead":"The paper tests a two-level stacking ensemble for brain age prediction, where the first level predicts age from voxels within each brain region and the second level combines those regional predictions. Across four MRI datasets, the ensemble beats the regional averaging baseline, reaching a mean absolute error of 4.75 years versus 5.68 years.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline MAE 4.75 conflates stacking with target-site adaptation: OOSPred S L1 P trains L0 on test-site labels, so the reported gain over GMV baselines is not purely attributable to the SE architecture.","rationale":"The paper's central claim, as stated in the abstract, is that the best SE setup reaches MAE 4.75 versus 5.68 for regional mean GMV. For that claim to be evidence for the proposed stacking ensemble, the improvement must be caused by the two-level ensemble rather than by the fact that the L0 stage in the winning setup sees the test site's age labels. The methods and results show exactly that confounding: OOSPred S L1 P obtains L0 predictions via 3-fold CV inside the test site, while the GMV ext baseline uses test-site labels only in L1. Comparing OOSPred S L1 P (4.75) with PredL0 S L1 P (5.12), which is the same SE design without test-site L0 training, attributes 0.37 years to target-site adaptation. The non-OOS SE-vs-GMV comparisons remain favorable, so the architecture may still be useful; the concern is about the headline number and the LOSO generalization framing. The biological insight section has a related circularity: L0 predictions are age-trained outputs, so their high correlation with age in Figures 4-6 is partly built into the model rather than discovered; however, the accuracy claim does not depend on that section. Overall the conditional verdict stands: the paper should add this control or clearly reframe the headline claim.","tokens_in":23146,"tokens_out":6713,"duration_ms":69119,"concrete_test":"Recompute the headline setup while holding L0 training strictly to the three LOSO training sites (no application-site labels), then train L1 on pooled training-site L0 predictions; compare the resulting MAE with the 4.75 of OOSPred S L1 P and the 5.12 of PredL0 S L1 P. As a stronger control, retrain L0 on test-site folds with shuffled age labels; if MAE stays near 4.75, adaptation is not the driver, whereas if it degrades, the headline gain is target-domain adaptation rather than stacking.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (MAE 4.75 vs 5.68) is obtained by the OOSPred S L1 P setup, in which L0 regional models are trained with 3-fold CV on the test site's own data, including its age labels (Methods, 'Data split and training setups'; Results). The GMV baseline compared against it, GMV P L1 P ext, uses test-site labels only at L1, not to construct the L0 features. The paper's non-OOS SE setup PredL0 S L1 P, whose L0 is trained only on the three training sites, achieves MAE 5.12. The 0.37-year gap between 5.12 and 4.75 is therefore attributable to using application-site labels at L0, not to the stacking architecture; and the entire 0.93-year gap over the ext baseline mixes this adaptation effect with the SE effect. The paper's claim that LOSO 'mimics the scenario when the models are applied to data from a new scanner not available during training' is thus invalid for the setup that produces the headline number. The broader claim that SE beats regional-mean GMV is not undermined, because non-OOS setups (e.g., PredL0 P L1 P MAE 5.06 vs GMV P L1 P 6.22) still show an advantage; but the abstract's strongest quantitative claim and the generalization narrative need to be reframed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-level stacking ensemble (SE) for MRI-based brain-age prediction: first-level (L0) models are trained on voxel-wise gray matter volume within each of 873 brain regions, and a second-level (L1) GLMnet model fuses the regional age predictions. Eight data-fusion scenarios are compared under leave-one-site-out (LOSO) validation across four datasets, with baselines using regional mean GMV. The authors report that SE outperforms the regional-mean baseline, with the abstract's headline result of MAE=4.75 for an out-of-sample L0 setup versus MAE=5.68 for the corresponding GMV baseline. They also report stronger age correlations of L0 predictions than of raw GMV, and lower dataset-identification accuracy from L0 predictions, which they interpret as improved biological insight and privacy.","tokens_in":23400,"tokens_out":6079,"duration_ms":59520,"significance":"If the central accuracy claim holds with the appropriate comparison, the paper offers a practically useful and privacy-aware framework for regional brain-age modeling, with an open code repository and a systematic exploration of data-fusion choices. The non-OOS comparisons (e.g., PredL0 P L1 P MAE=5.06 vs GMV P L1 P MAE=6.22) do show an advantage for the stacking architecture over regional-mean GMV, so the qualitative conclusion is defensible. The paper is less convincing on its headline framing, statistical substantiation, and biological-interpretation claims.","major_comments":[{"comment":"The abstract's headline comparison (MAE=4.75 vs 5.68) conflates the stacking ensemble with target-domain adaptation. In OOSPred S L1 P, the L0 regional models are trained with 3-fold CV on the test site's own data, including its age labels, while the GMV P L1 P ext baseline only uses test-site folds at L1 and a fixed averaging operation at L0. The paper's statement that LOSO 'mimics the scenario when the models are applied to data from a new scanner not available during training' is therefore invalid for the setup that produces the headline number. The non-OOS setups (e.g., PredL0 S L1 P, MAE=5.12, vs GMV P L1 P, MAE=6.22) preserve the qualitative SE advantage, but the abstract and Results need to be reframed so the 4.75 result is attributed to SE combined with test-site L0 fitting, not to SE alone.","section":"Data split and training setups; Results; Supplementary Table 2"},{"comment":"The claim that L0 predictions 'better capture the aging process' is circular: the L0 models are trained to predict age, so their outputs are mechanically correlated with age. The comparison of L0 prediction-age correlations with raw GMV-age correlations therefore does not provide independent evidence of improved biological validity. The paper should either remove or substantially soften this interpretation, or support it with held-out validation, external age-related outcomes, or a non-circular benchmark such as predicting age from GMV in one region using a model trained on the same region's GMV (rather than on age itself).","section":"Results, Biological insights; Figures 4-6"},{"comment":"No uncertainty quantification is provided for the LOSO performance differences. With only four test sites, each setup yields a single aggregate MAE; the reported differences (e.g., 4.75 vs 5.12 vs 5.68) could be within site-to-site noise. The paper should report per-fold MAE values, confidence intervals, or a paired statistical test across the four LOSO folds before claiming a reliable ranking of setups. This is load-bearing for the central claim that SE outperforms the GMV baselines.","section":"Results, Performance; Figure 2; Supplementary Table 2"}],"minor_comments":[{"comment":"The expression 'p<<0.01e13' is not a valid p-value; please correct the notation and report a test statistic or a standard p-value.","section":"Results, Biological insights"},{"comment":"The two sets of mean correlations (r_mean=0.6 vs 0.32 in the Biological insights section and mean=0.86 vs 0.81 in the Stability across datasets section) appear inconsistent; the text should clarify which quantity is averaged and how the two analyses differ.","section":"Results, Biological insights vs Stability across datasets"},{"comment":"The duplicate rows marked with an asterisk and the notation 'Mean()' in the L0 operation column are confusing; please clarify whether the GMV setups use the same training splits and data as the corresponding OOS setups.","section":"Table 1"},{"comment":"The description of GLMnet's 'adaptive resampling' for tuning lambda and alpha is brief; for reproducibility, please specify the resampling scheme (number of folds, repetitions) or cite the package defaults explicitly.","section":"Methods, SE model training"},{"comment":"The sentence 'This way of obtaining OOS L0 predictions effectively models the idiosyncrasies of the specific dataset' is vague and risks overinterpreting a target-domain adaptation effect; please revise to state precisely what is being modeled and how.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper's central methodological contribution is potentially sound, and the code availability is a clear strength. The main concerns are the misleading headline comparison and the circular biological-interpretation claim; both are addressable by reframing and additional analyses. I would ask the authors to report per-fold results and to add a comparison with a GMV-based L0 model trained on test-site folds if they wish to keep the 4.75 result. The paper is within scope for a neuroimaging methods journal, and the declared conflict of interest is appropriately disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Region-wise stacking for brain age is a real contribution: the paper formulates a two-level ensemble where L0 models are per-region voxel-based age predictors and L1 fuses them, and it explores eight fusion scenarios systematically. The comparison against the standard regional-mean GMV baseline looks fair for the non-OOS setups, and those still show an advantage (e.g., PredL0 P L1 P at MAE 5.06 vs GMV P L1 P at 6.22). Code is public, the datasets are standard, and the paper is honest that it does not compare against SOTA.\n\nThe main weak spot is the headline: the best setup, OOSPred S L1 P (MAE 4.75 vs 5.68), obtains L0 predictions by running 3-fold CV on the test site itself, using its age labels to train the first-level models. That gain over the baseline is therefore largely target-domain adaptation, not a property of the stacking architecture. The paper's statement that LOSO mimics a new-scanner scenario is only valid for the non-OOS setups. The stress-test note is right, and this should be reframed: report the non-OOS SE numbers as the architecture's contribution, and treat OOS as a separate privacy-preserving workflow.\n\nThe biological insight section is also circular. L0 models are trained to predict age, so high correlation between L0 outputs and age is expected; presenting it as evidence that SE 'better captures the aging process' does not hold. The more defensible point is that L0 predictions retain age information per region, which is what stacking needs.\n\nSmaller issues: no confidence intervals or significance tests for MAE differences across the four LOSO folds; the privacy analysis is a reasonable proxy but not a guarantee; bias results are mixed and not discussed deeply. None of this sinks the core claim. The central comparison is clean enough, and the utility for multi-site brain age work is plausible.\n\nWho is this for: neuroimaging method developers and anyone using regional GMV features for age prediction. It deserves a serious referee; I would accept with revisions, mainly reframing the headline and the biological claims.","headline":"Region-wise stacking is a genuine contribution and the core accuracy claim holds up outside the OOS setups, but the headline number conflates stacking with target-domain adaptation.","tokens_in":23955,"tokens_out":1968,"would_cite":true,"duration_ms":19412,"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":"A two-level stacking ensemble, pairing per-region voxel-based age predictors with a second-stage fusion model, estimates brain age from structural MRI more accurately than does the conventional regional-mean gray-matter baseline.","keywords":["brain-age prediction","stacking ensemble","magnetic resonance imaging","gray matter volume","voxel-based morphometry","data fusion","leave-one-site-out validation","data privacy"],"falsifier":"Train the first-level regional models strictly on the three held-in sites, with no target-site data, and compare the same mean absolute error against the regional-mean gray-matter baseline; if the stacking gain shrinks or reverses, the claimed superiority over averaging would be refuted.","tokens_in":22947,"feed_emoji":"🧠","tokens_out":11796,"duration_ms":100520,"temperature":0.7,"pith_summary":"The paper proposes a two-level stacking ensemble for estimating a person's age from structural MRI. Instead of averaging gray-matter volume across each brain region, the first level trains a separate voxel-based predictor of age for each of 873 regions, and the second level learns how to combine those 873 regional age predictions into one estimate. Across four datasets and multiple fusion setups, the stacking ensembles beat the classical regional-mean baseline; the best setup reaches a mean absolute error of 4.75 years versus 5.68 years for the baseline. The paper also reports that regional age predictions carry a stronger and more consistent aging signal than regional volume averages, and that they make it harder to identify which dataset a scan came from, which it interprets as a privacy benefit. The intended significance is a more accurate, interpretable, and privacy-preserving route to brain-age biomarkers for studying healthy and pathological aging.","feed_headline":"Two-level stacking predicts brain age from MRI to 4.75 years","feed_subtitle":"Replacing regional GMV averaging with stacked regional age models improves accuracy and privacy across four datasets.","key_machinery":"The central object is the two-level stacking ensemble. At level 0, each of 873 predefined brain regions gets its own elastic-net regression model, trained on the voxel-wise gray matter volume values inside that region to predict age; this replaces the uniform averaging of voxels used in the baseline. At level 1, a second elastic-net model takes the 873 regional age predictions as input features and produces the final age estimate. The mechanism works because the first-level models weight each voxel by how well it tracks aging instead of treating all voxels equally, and the second-level model learns how much to trust each region rather than applying a predefined weight. The paper also relies on an out-of-sample scheme for the best setup: L0 predictions for a target site are obtained by three-fold cross-validation on that site's own data, with L1 trained on independent data from other sites.","core_discovery":"On its own terms, the paper's central claim is that a two-level stacking ensemble improves brain-age prediction compared with the standard practice of averaging voxel-wise gray matter volume within brain regions. At the first level (L0), an elastic-net model is trained per region on the voxels of that region to predict age; at the second level (L1), another elastic-net model learns to combine the 873 regional age predictions. The paper systematically compares eight data-fusion scenarios across four datasets with leave-one-site-out validation. The best-performing scenario obtains out-of-sample L0 predictions at the application site and trains L1 on independent site-specific data, achieving a mean absolute error of 4.75 years versus 5.68 years for the regional-mean baseline; all stacking setups outperform the gray-matter-volume baselines on mean absolute error, $\\mathrm{R}^2$, and correlation. The paper further claims that L0 predictions correlate more strongly and more stably with chronological age than regional mean gray matter volume (mean absolute correlation 0.6 versus 0.32), and that a dataset-of-origin classifier achieves lower balanced accuracy on L0 predictions (0.63) than on gray-matter-volume features (0.87), indicating improved privacy.","pith_inferences":["Because the best out-of-sample setup trains L0 models on the target site's own data with known ages, part of the reported gain may come from target-site adaptation rather than from the stacking architecture; a stricter evaluation would train L0 exclusively on the three held-in sites.","The same two-level design could be applied to other voxel-wise biomarkers or to longitudinal settings, and age-bias correction could be integrated into the L1 model rather than applied post hoc.","A direct ablation comparing the stacking ensemble with a single whole-brain voxel-based elastic net under the same regularization and folds would isolate whether the regional decomposition and the second-level fusion carry the gain.","The privacy conclusion rests on a proxy task; a direct re-identification test would strengthen or weaken the claim that L0 predictions protect subject identity."],"forward_implications":["A clinic or consortium can estimate brain age without pooling raw MRI scans, because only the 873 regional age predictions need to be shared.","Regional age predictions reveal aging signal in subcortical structures that regional volume averaging dilutes, giving a more detailed map of where the brain ages.","Accuracy improves systematically as more training datasets are added, so the approach becomes more valuable as multi-site data accumulate.","The two-level design leaves room for other base learners, parcellation schemes, and bias-correction steps, which the paper identifies as future refinements."],"supporting_citations":[{"why":"Supplies the stacking-generalization method that the two-level ensemble is built on.","marker":"[38]"},{"why":"Provides the elastic-net regression used at both the first and second levels.","marker":"[47]"},{"why":"Defines the 873-region parcellation that groups voxels for the first-level regional models.","marker":"[31]"},{"why":"Supplies one of the four adult-lifespan datasets used for training and testing.","marker":"[43]"},{"why":"Supplies one of the four adult-lifespan datasets used for training and testing.","marker":"[44]"},{"why":"Supplies one of the four adult-lifespan datasets used for training and testing.","marker":"[45]"}],"fun_headline_variants":["Stacked regional models sharpen brain-age MRI predictions","Two-level stacking beats voxel averaging for brain-age","Brain-age MRI error drops to 4.75 years with stacking","Regional stacking trims brain-age error and boosts privacy","Stacking MRI regions improves age prediction accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that stacking beats regional averaging stands on the assumption that the accuracy gain comes from the stacking design itself, not from the best-performing setup's use of the target site's own data, with known ages, to train the first-level models.","fun_headline_variants_meta":{"raw":{"variants":["Stacked regional models sharpen brain-age MRI predictions","Two-level stacking beats voxel averaging for brain-age","Brain-age MRI error drops to 4.75 years with stacking","Regional stacking trims brain-age error and boosts privacy","Stacking MRI regions improves age prediction accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00083,"raw_usage":{"total_tokens":3690,"prompt_tokens":1076,"completion_tokens":2614,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":2520}},"tokens_in":692,"tokens_out":2614,"duration_ms":20651,"temperature":1.0,"reasoning_tokens":2520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:22:38.793393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the first-level regional models strictly on the three held-in sites, with no target-site data, and compare the same mean absolute error against the regional-mean gray-matter baseline; if the stacking gain shrinks or reverses, the claimed superiority over averaging would be refuted.","supporting_citations":[{"cited_title":"Advanced brain ageing in Parkinson’s disease is related to disease duration and individual impairment,","cited_arxiv_id":null,"evidence_quote":"Defines the 873-region parcellation that groups voxels for the first-level regional models."},{"cited_title":"Studying variability in human brain aging in a population-based German cohort—rationale and design of 1000BRAINS,","cited_arxiv_id":null,"evidence_quote":"Supplies one of the four adult-lifespan datasets used for training and testing."}],"review_version":1}