{"id":"b5a554b4-03d0-4ca5-81f8-cbca001dacf1","arxiv_id":"2502.08070","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cerebral blood flow peaks at about 7.1 years and then declines, and normative charts from over 9,000 scans flag disease-specific perfusion deviations.","lead":"Researchers pooled brain blood flow scans from 20 datasets to make growth charts for cerebral perfusion from birth to age 85, and found blood flow peaks around age 7 before declining. The charts let doctors see when an individual's brain blood flow is unusually low, with early tests in Alzheimer's, dementia, and depression patients.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The normative peak age and adult decline rest on an untested assumption that site and sequence effects are age-invariant; a sequence-age interaction could shift the estimated peak outside the reported CI.","rationale":"The paper has genuine strengths: a large aggregated sample, explicit QEI quality control, GAMLSS with BIC-based distribution selection, and multiple sensitivity analyses including LOSO and balanced resampling. Those analyses demonstrate that no single site or sparse age bin dominates the result, but they do not resolve the more fundamental identification problem: the site and sequence adjustments in Eqs. 3-4 assume age-invariant additive effects. Because ASL quantification biases can vary with age through arterial transit time, labeling efficiency, and partial-volume effects, the normative curve's shape—especially the sharp rise to a 7.1-year peak—could be shaped by sequence-age confounding. The reader's weakest assumption identifies exactly this risk, and the Discussion's admission that sequence effects are not fully mitigated makes the concern explicit in the manuscript itself. My independent reading does not move the verdict: CONDITIONAL remains appropriate, with the proposed interaction test as the natural condition for accepting the trajectory as biological rather than methodological.","tokens_in":24225,"tokens_out":3823,"duration_ms":36307,"concrete_test":"Refit the global CBF GAMLSS model (Eq. 3) adding an age×sequence interaction term in μ, and separately refit within each major sequence family (e.g., PASL vs PCASL). If the estimated peak age moves outside the reported 95% CI of 6.5-7.5 years, or the adult decline slope changes by more than 20%, the sequence-age additivity assumption fails and the normative trajectory is not settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the lifespan CBF trajectory with peak at 7.1 years (95% CI: 6.5-7.5) and a gradual near-linear adult decline. Equations 3-4 model μ(age) with a B-spline plus sex and sequence fixed effects and site as a random intercept. This identifies the age trajectory only if site and sequence effects are additive and constant across age. But pediatric and adult cohorts differ systematically in ASL sequence (PASL vs PCASL, single vs multi-PLD, M0 handling), and these differences can produce age-dependent CBF biases: arterial transit time changes with age, so labeling efficiency and PLD assumptions affect children and adults differently. A fixed sequence intercept shifts all ages equally; it cannot absorb a sequence×age interaction or site×age shape differences. If such interactions exist, the sharp postnatal rise and the 7.1-year peak are confounded: no within-site longitudinal curve is presented to support the rise, and the adult decline could partly reflect older cohorts scanned with different sequences rather than biological aging. The LOSO and balanced-resampling analyses reduce leverage of individual sites and sparse age bins, but they do not test the additivity assumption itself. The Discussion explicitly concedes that sequence factors are not fully mitigated. Therefore the most load-bearing unsecured condition is age-invariance of sequence/site effects, and it needs a direct test before the normative trajectory can be treated as biological.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper assembles a large multi-site ASL dataset and constructs normative models of cerebral perfusion across ages 0-85 using GAMLSS with a GB2 distribution. After quality control, the sample includes 8,460 healthy controls and 903 patients from 20 datasets. The central results are a global CBF peak at approximately 7.1 years (95% CI 6.5-7.5) followed by a near-linear adult decline, region-specific peak ages, sex differences, and normative deviation scores applied to AD, FTD, MCI, and MDD. Sensitivity analyses include stricter QEI thresholds, balanced resampling, split-half replication, bootstrap, and leave-one-site-out. The paper also reports longitudinal progression trajectories in AD and MCI, including MCI-A and MCI-N subtypes.","tokens_in":24545,"tokens_out":5806,"duration_ms":52590,"significance":"If the normative trajectory is unbiased, this would be a valuable resource analogous to structural brain charts, providing age- and sex-adjusted CBF norms and individualized deviation scores. The scale of the healthy-control sample, the explicit external-split structure for clinical applications, and the battery of sensitivity analyses are genuine strengths. The longitudinal MCI subtype analysis uses external outcomes (conversion, cognitive decline) and is a useful demonstration of prognostic value. However, the central age trajectory rests on an untested additivity assumption for site and ASL sequence effects, and the claimed birth-to-85 coverage is stronger than the acknowledged sparse sampling supports. The clinical significance is therefore conditional on the resolution of these methodological points.","major_comments":[{"comment":"The model identifies the age trajectory only if site and ASL sequence effects are age-invariant. The fixed sequence term and the random site intercept in Eqs. (3)-(4) shift the mean across all ages equally and cannot absorb sequence-by-age or site-by-age interactions. Because pediatric and adult cohorts differ systematically in ASL sequence type, PLD, and M0 handling, and because arterial transit time changes with age, such interactions would bias the shape of the curve and the estimated peak age. The LOSO and balanced-resampling analyses do not test the additivity assumption itself. Please provide a direct test, for example by including an age-by-sequence interaction or site-specific age terms and reporting how the peak age and decline rate change, or by presenting within-site age trajectories for sites with wide age coverage.","section":"Methods, Eqs. (3)-(4); Discussion, first limitation"},{"comment":"The claimed lifespan range 'birth to 85 years' overstates the empirical support because ages 0-5 and 30-40 are acknowledged to be underrepresented. Balanced resampling to 64 participants per five-year bin does not add information in sparse bins; it only equalizes weights. Please report the number of healthy participants in each five-year age bin (especially 0-5) in the main text or a main-table version of Supplementary Table 2, and rerun the global model excluding the sparse bins to quantify how dependent the 7.1-year peak and the adult decline are on those observations.","section":"Discussion, fourth limitation; Methods, balanced resampling"},{"comment":"The baseline TNP difference between the MCI-A and MCI-N subtypes is partly expected by construction, because the subtypes are obtained by k-means clustering on the same regional and network deviation scores used to compute TNP. The statement that 'Baseline TNP was significantly higher in MCI-A compared to MCI-N' is therefore a clustering check rather than an independent clinical finding. The independent evidence is the difference in longitudinal cognitive decline and conversion risk, which should be emphasized as the primary support for subtype validity; the baseline TNP contrast should be reframed accordingly.","section":"Results, Figure 3E; Methods, MCI subtype identification"},{"comment":"The abstract states that the model identifies 'disease-specific perfusion abnormalities across four brain disorders,' but the MDD classification performance is not significant (mean AUC = 54.88, p = 0.272). The post hoc explanation of imbalanced sample size is plausible but is not demonstrated. Please either soften the abstract and general claims to three disorders with significant classification, or provide a matched-control analysis showing that the MDD null result is attributable to sample imbalance.","section":"Results, Figure 3F; Abstract"}],"minor_comments":[{"comment":"The abstract says 'over 12,000 high-quality ASL MRI scans,' while the Results report 9,363 participants after QC. Clarify whether these are scans or participants, and reconcile the numbers.","section":"Abstract vs. Results"},{"comment":"Equation (2) uses the symbol gamma for the response variable, which is inconsistent with the text describing CBF values as y. This appears to be a typographical issue.","section":"Methods, Eq. (2)"},{"comment":"The sentence 'Because ASL MRI does need any exogenous contrast agents' should read 'does not need.'","section":"Introduction"},{"comment":"The atlas description states 48 cortical and 8 subcortical regions, while the longitudinal TNP definition uses '58 cortical and 8 subcortical regions.' Since TNP is a proportion, the denominator must be consistent; please correct this discrepancy.","section":"Methods, atlas regions; Methods, longitudinal TNP"},{"comment":"The sentence 'Prevent AD and Dalas lifespan were set as reference health control dataset... the ADNI were set as disease dataset, included 649 health controls with 213 scans' contains likely typos ('Dalas' and the numbers 649/213), and the dataset names should be spelled consistently.","section":"Methods, longitudinal dataset description"},{"comment":"Figure 1B is labeled as sex-specific trajectories, but Eq. (3) includes sex only as an additive fixed effect, which would produce parallel curves. If sex-by-age interactions were not modeled, the figure legend should clarify that sex differences are constant offsets.","section":"Figure 1B and Methods, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially useful resource paper with a large and carefully processed multi-site ASL sample. My main reservation is the untested age-invariance of site and sequence effects, which directly affects the peak-age claim that is the headline result; the sparse age coverage compounds this concern. The clinical sections are mostly secondary, but the MDD null result should not be presented as supporting 'four brain disorders' without qualification. If the authors can provide a direct interaction test, report age-bin counts, and tighten the overclaims, I would be willing to reconsider."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a real resource paper. The genuinely new thing is not that CBF peaks around age 7 and declines afterward—that shape has been reported before—but that they actually built the charts at global, network, and regional levels and turned them into individual deviation scores. That is useful for ASL neuroimaging in the same way the Bethlehem brain charts are useful for structural imaging, and it is the first time GAMLSS normative modeling has been applied to ASL perfusion with this kind of scale.\n\nWhat I like: 8,460 healthy controls pooled from 20 datasets, careful QEI-based quality control, GB2 distribution selection, and a serious battery of sensitivity analyses—QEI threshold change, balanced resampling, split-half, bootstrap, and leave-one-site-out. The patient application is a bonus, not the core, and the longitudinal MCI progression analysis is a reasonable demonstration of how the charts could be used clinically. The paper earns credit for shipping this much empirical work as a preprint.\n\nThe soft spot is the one the stress-test note names: the model identifies the age trajectory by assuming sequence and site effects are additive and age-invariant. Pediatric and adult cohorts differ systematically in ASL sequence, and arterial transit time changes with age, so a fixed sequence intercept cannot absorb sequence-by-age bias. LOSO and balanced resampling reduce the leverage of any single site or sparse age bin, but they do not test the additivity assumption itself. The Discussion explicitly admits that sequence factors are not fully mitigated. I do not think this collapses the paper: the broad trajectory is consistent with prior ASL and PET literature, and the sensitivity analyses are reassuring. But the 95% CI on the 7.1-year peak is probably too tight, and the claim of covering birth to 85 years overstates the sparse 0–5 and 30–40 bins.\n\nThe MCI subtype clustering is partly circular because the subtypes come from clustering the same deviation scores that define them, but the longitudinal divergence in MMSE, CDR, and conversion odds is a genuine out-of-sample signal, so I would call that a minor issue rather than a fatal one. What bothers me more for a resource paper is the absence of code and data artifacts; a normative chart paper should ship the model parameters or at least a public implementation.\n\nBottom line: this deserves a serious referee. I would send it out and ask for code/data release, an explicit age-coverage caveat, and a direct test of the additivity assumption—for example, within-sequence age curves or a sequence-by-age interaction term. If those hold, this is a citeable resource.","headline":"A potentially citeable normative CBF resource that deserves a serious referee; the peak-age estimate rests on an untested sequence-by-age additivity assumption that the Discussion itself concedes.","tokens_in":25078,"tokens_out":1666,"would_cite":true,"duration_ms":16632,"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":"Brain blood flow peaks around age 7 and then declines for life, and deviations from this curve can flag disease-specific perfusion abnormalities.","keywords":["cerebral blood flow","arterial spin labeling MRI","normative modeling","lifespan brain trajectories","GAMLSS","Alzheimer's disease","mild cognitive impairment","neurodevelopment"],"falsifier":"One could re-fit the GAMLSS models while allowing age-by-sequence-type interaction terms and recompute the global peak; if the peak moves outside the 6.5–7.5 year interval, the 7.1-year peak is an artifact of pooling. A complementary check is to estimate the same curves using only sites with a single uniform ASL sequence and compare the resulting peak age and slope.","tokens_in":24047,"feed_emoji":"🧠","tokens_out":10987,"duration_ms":82982,"temperature":0.7,"pith_summary":"This paper sets out to give cerebral perfusion—the rate at which blood delivers oxygen and nutrients to brain tissue—a normative chart covering the entire human lifespan, from birth to 85 years. Using arterial spin labeling (ASL) MRI scans from 12,633 participants, of whom 9,363 passed quality control, the authors model cerebral blood flow as a smooth function of age. The resulting trajectories show a rapid postnatal rise, a global peak at approximately 7.1 years (95% CI: 6.5–7.5), and a gradual near-linear decline into old age, with regional peaks ranging from 6.2 to 8.6 years. The paper then converts each person's position on these curves into a deviation score and uses those scores to detect disease-specific perfusion abnormalities in Alzheimer's disease, frontotemporal dementia, mild cognitive impairment, and major depressive disorder. If the model holds, perfusion becomes a quantitative, age- and sex-adjusted biomarker that could flag abnormal brain development or neurodegeneration from a single scan.","feed_headline":"Brain blood flow peaks at 7.1 years, chart shows","feed_subtitle":"Normative curves from 9,363 brain scans flag Alzheimer's, FTD, and MCI deviations.","key_machinery":"The machinery is distributional regression using a generalized additive model for location, scale, and shape, applied with a four-parameter Generalized Beta 2 distribution. The median and the coefficient of variation are each modeled as smooth B-spline functions of age, with sex and ASL sequence type as fixed effects and dataset site as a random effect, so the model captures both the mean growth curve and the way between-person variability changes across life. Individual deviation scores are computed by converting a person's fitted GB2 quantile into a standard normal z-score; extreme negative deviations (z < −2.3) are counted regionally and used to define MCI subtypes, classify disorders, and model longitudinal progression. Peak ages come from the first derivative of the modeled median curve crossing zero, and all curves are validated with bootstrap, split-half, balanced-resampling, and leave-one-site-out analyses.","core_discovery":"The central discovery is a normative, age- and sex-aware lifespan trajectory for human cerebral blood flow: global CBF rises sharply after birth, peaks at about 7.1 years (95% CI: 6.5–7.5), and then declines gradually and near-linearly through adulthood. Cortical perfusion peaks highest in childhood; white matter peaks earlier (about 5.5 years) and declines more slowly; subcortical perfusion peaks around 6.7 years. Females show significantly higher perfusion than males in global, cortical, white matter, and subcortical measures, with a significant variance difference only in white matter. Individual deviation z-scores derived from the model reveal disease-specific hypoperfusion patterns, and classification analyses separate frontotemporal dementia (mean AUC 87.67) and Alzheimer's disease (mean AUC 80.24) from controls. In mild cognitive impairment, an AD-like perfusion subtype shows faster cognitive decline and higher odds of conversion to AD than a normal-perfusion subtype.","pith_inferences":["Editorial inference: the pooling design can be stress-tested by adding age-by-sequence interaction terms; if the global peak then moves outside the reported 6.5–7.5 year interval, the 7.1-year peak would be a harmonization artifact rather than a biological milestone.","Editorial inference: the regional maturation gradient implies a testable prediction outside the paper—children with later prefrontal perfusion peaks should show later maturation of executive functions.","Editorial inference: the high between-subject perfusion variability seen in childhood suggests that deviation scores early in life may be less reliable than adult scores, so clinical use in children would require age-specific reliability estimates.","Editorial inference: the same normative machinery could be extended to other ASL-derived physiological measures, such as arterial transit time or blood-brain-barrier water exchange, which the paper notes were not available in most of its datasets."],"forward_implications":["If the model is correct, a single ASL scan can be compared against an age- and sex-matched reference, allowing a child with abnormally low perfusion or an adult with accelerated decline to be identified against a quantitative standard.","Perfusion milestones can be aligned with structural milestones, such as global perfusion peaking at 7.1 years and gray matter volume peaking at 7.8 years, giving developmental and aging research a shared temporal benchmark.","Perfusion deviation scores provide a transdiagnostic measure: AD and FTD separate from healthy controls with mean AUCs of 80.24 and 87.67, while MCI divides into AD-like and normal-perfusion subtypes with different conversion rates.","Longitudinal modeling shows that the proportion of extreme negative perfusion deviations increases faster in AD than in controls and faster in progressive than in stable MCI, supporting perfusion-based monitoring of disease progression.","Because sequence type and site are explicit terms in the model, the normative chart can in principle be reused across heterogeneous multi-site datasets, although the paper acknowledges sequence-related variation as a limitation."],"supporting_citations":[{"why":"Supplies the normative brain-chart approach and the GAMLSS growth-curve methodology that this paper extends to cerebral perfusion, plus the gray matter volume peak used as a milestone comparison.","marker":"Bethlehem et al. (2022)"},{"why":"Defines the GAMLSS statistical framework used to model age trajectories of the median, variance, skewness, and kurtosis of CBF.","marker":"Stasinopoulos & Rigby (2008)"},{"why":"Establishes the growth-standard methodology for constructing normative growth curves.","marker":"Borghi et al. (2006)"},{"why":"Provides the automated quality evaluation index (QEI) used to exclude low-quality ASL scans before modeling.","marker":"Dolui et al. (2024)"},{"why":"Sets out the normative modeling protocol for computing and validating individual deviation z-scores.","marker":"Rutherford et al. (2022)"},{"why":"Provides the conceptual foundation for treating brain disorders as deviations from normative functioning.","marker":"Marquand et al. (2019)"},{"why":"Supplies the longitudinal normative modeling approach and the extreme-deviation threshold convention used for disease progression.","marker":"Verdi et al. (2024)"},{"why":"Describes the ASL processing toolbox used to quantify CBF maps in physiological units.","marker":"Wang, Aguirre, et al. (2008)"},{"why":"Supplies the cortical and subcortical atlas used to define regional CBF measures.","marker":"Desikan et al. (2006)"},{"why":"Supplies the network atlas used to compute network-level CBF values.","marker":"Yeo et al. (2011)"}],"fun_headline_variants":["Brain blood flow peaks at 7.1 years, then declines","Normative brain perfusion peaking at 7.1, then falling","Peak cerebral perfusion at 7.1 years, then gradual drop","Charting brain blood flow: peak at 7.1, decline after","Brain perfusion tops at 7.1 years, then slides down"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes that pooling cross-sectional ASL scans from 20 sites with different acquisition sequences, and then adjusting for sequence type as a fixed effect and site as a random effect, yields unbiased age trajectories; if site or sequence effects interact with age, the estimated peak age of 7.1 years, the decline shape, and all patient deviation scores would be biased.","fun_headline_variants_meta":{"raw":{"variants":["Brain blood flow peaks at 7.1 years, then declines","Normative brain perfusion peaking at 7.1, then falling","Peak cerebral perfusion at 7.1 years, then gradual drop","Charting brain blood flow: peak at 7.1, decline after","Brain perfusion tops at 7.1 years, then slides down"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000813,"raw_usage":{"total_tokens":3582,"prompt_tokens":982,"completion_tokens":2600,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2505}},"tokens_in":598,"tokens_out":2600,"duration_ms":46503,"temperature":1.0,"reasoning_tokens":2505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:56:06.014353+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One could re-fit the GAMLSS models while allowing age-by-sequence-type interaction terms and recompute the global peak; if the peak moves outside the 6.5–7.5 year interval, the 7.1-year peak is an artifact of pooling. A complementary check is to estimate the same curves using only sites with a single uniform ASL sequence and compare the resulting peak age and slope.","supporting_citations":[],"review_version":1}