pith. sign in

arxiv: 2605.14565 · v2 · pith:VNOIYJJLnew · submitted 2026-05-14 · 📊 stat.ME · math.ST· stat.AP· stat.TH

A Bayesian Longitudinal Spatial Normative Model for Individualized Brain Deviation Mapping

J. T. Korley This is my paper

Pith reviewed 2026-05-20 21:30 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.APstat.TH
keywords Bayesian normative modelinglongitudinal spatial statisticsbrain deviation mappinghierarchical modelingstructural MRIAlzheimer neurodegenerationindividualized mappinglatent spatial process
0
0 comments X

The pith

A Bayesian model jointly captures time trends and spatial brain patterns to map individual deviations more accurately.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a statistical model that processes repeated brain scans from the same people while respecting how deviations in one brain area relate to nearby areas. It builds a single hierarchical setup that estimates each person's deviation map as a hidden spatial field whose best guess comes directly from the posterior distribution. A reader would care because standard approaches either treat regions in isolation or ignore repeated measures, which inflates error when trying to spot personal departures from normal aging. The new estimator is shown to lower reconstruction error in simulations that vary spatial strength, trajectory shape, visit timing, and missing data. On real OASIS-3 MRI scans the model cuts RMSE by 54 percent versus a cross-sectional baseline and by 45 percent versus a longitudinal but non-spatial baseline, with deviations clustering in temporal and cingulate regions.

Core claim

The Bayesian longitudinal spatial normative model jointly captures within-subject temporal dependence and spatially structured subject-specific deviations within a unified hierarchical framework. The individualized deviation map is treated as a latent spatial process with an explicit posterior distribution, yielding a principled Bayes estimator under squared error loss rather than an ad hoc residual summary. Across six simulation scenarios the model reduces reconstruction error while remaining well calibrated; in OASIS-3 data it lowers RMSE by 54 percent relative to independent cross-sectional models and by 45 percent relative to longitudinal non-spatial models, with regional burden highest,

What carries the argument

The latent spatial process representation of the individualized deviation map inside the unified hierarchical Bayesian framework that couples temporal and spatial dependence.

If this is right

  • The model reduces deviation-map reconstruction error across simulations that include nonlinear trajectories, irregular visit schedules, and missing follow-ups.
  • On OASIS-3 structural MRI the approach produces 54 percent lower RMSE than independent cross-sectional models and 45 percent lower RMSE than longitudinal non-spatial models.
  • Regional deviation burden concentrates in the temporal pole, entorhinal cortex, inferior temporal cortex, posterior cingulate, and parahippocampal cortex.
  • Subject-level deviation profiles exhibit marked heterogeneity even when global cognitive scores remain comparable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same latent-process construction could be reused to incorporate functional or diffusion imaging for multi-modal deviation maps.
  • The resulting subject-specific profiles might serve as more sensitive endpoints in clinical trials that track early neurodegeneration.
  • Testing whether the deviation maps predict subsequent cognitive decline on held-out visits would directly link the estimator to clinical utility.
  • Adding genetic or vascular covariates into the normative reference layer could further tighten the spatial covariance estimates.

Load-bearing premise

The spatial organization of neuroanatomical deviations can be captured by a covariance structure placed on the latent deviation process.

What would settle it

A new longitudinal MRI dataset that supplies independent ground-truth deviation maps; if the proposed model's posterior mean then shows higher squared error than the two benchmark estimators, the claimed advantage disappears.

Figures

Figures reproduced from arXiv: 2605.14565 by J. T. Korley.

Figure 3
Figure 3. Figure 3: Model comparison in the OASIS-3 real-data application. The Bayesian subject-specific spatial model pro [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Estimation accuracy and deviation-map reconstruction across simulation scenarios. [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cortical burden of extreme standardized deviations in the OASIS-3 application. The map displays [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 2
Figure 2. Figure 2: Calibration under varying spatial and longitudinal data-generating mechanisms. Dashed [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Longitudinal standardized deviation trajectories for selected structural regions. Gray lines represent subject [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 3
Figure 3. Figure 3: Model comparison in the OASIS-3 real-data application. The Bayesian subject-specific [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustrative subject-level deviation profiles. The upper panel shows the subject with the largest mean absolute [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cortical burden of extreme standardized deviations in the OASIS-3 application. The [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Longitudinal standardized deviation trajectories for selected structural regions. Gray [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustrative subject-level deviation profiles. The upper panel shows the subject with [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
read the original abstract

Normative modeling enables individualized characterization of structural brain deviations by evaluating subjects against a reference population rather than a group average. Most existing implementations treat brain regions independently and remain cross-sectional, despite the availability of repeated neuroimaging measurements and the well-documented spatial organization of neuroanatomical variation. We propose a Bayesian longitudinal spatial normative model that jointly captures within-subject temporal dependence and spatially structured subject-specific deviations within a unified hierarchical framework. The individualized deviation map is treated as a latent spatial process with an explicit posterior distribution, yielding a principled Bayes estimator under squared error loss rather than an ad hoc residual summary. Across six simulation scenarios encompassing varying spatial dependence, nonlinear trajectories, irregular visit schedules, and missing follow-up, the proposed model consistently reduced deviation-map reconstruction error relative to independent cross-sectional and longitudinal non-spatial benchmarks while maintaining stable calibration. In an application to OASIS-3 structural MRI data, the model reduced RMSE by 54% relative to the independent cross-sectional model and by 45% relative to the longitudinal non-spatial model. Regional deviation burden was concentrated in the temporal pole, entorhinal cortex, inferior temporal cortex, posterior cingulate, and parahippocampal cortex, consistent with regions implicated in early Alzheimer-type neurodegeneration. Subject-level profiles revealed substantial heterogeneity in regional abnormality patterns, including marked multiregional deviation with preserved global cognitive scores.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a Bayesian hierarchical normative model for individualized brain deviation mapping from longitudinal structural MRI. It jointly models within-subject temporal dependence and spatially structured subject-specific deviations by treating the deviation map as a latent spatial process with an explicit posterior distribution, yielding the posterior mean as a Bayes estimator under squared-error loss. The model is tested across six simulation scenarios with varying spatial dependence, nonlinear trajectories, irregular visits, and missing data, and applied to OASIS-3 data where it reports 54% RMSE reduction versus an independent cross-sectional benchmark and 45% versus a longitudinal non-spatial model. Regional deviation burden is concentrated in temporal pole, entorhinal, inferior temporal, posterior cingulate, and parahippocampal cortex, with subject-level heterogeneity noted.

Significance. If the central performance claims hold under the stated modeling assumptions, the work offers a statistically coherent advance over existing normative approaches by unifying longitudinal and spatial structure in a single hierarchical posterior rather than ad-hoc residuals. The explicit treatment of the deviation map as a latent spatial process and the consistent error reductions across controlled simulations constitute clear strengths. Successful validation could improve sensitivity for early neurodegenerative patterns while preserving subject-specific profiles.

major comments (2)
  1. [Model specification / Results (OASIS-3 application)] The central performance gains (RMSE reductions of 54% and 45%) rest on the assumption that the chosen spatial covariance for the latent deviation process adequately captures neuroanatomical correlation structure in the OASIS-3 parcellation. Without a sensitivity analysis to alternative kernels or graph Laplacians, or a direct comparison of fitted versus empirical spatial correlations, it remains possible that gains arise from added flexibility rather than correct dependence modeling.
  2. [Methods / Simulation study] The abstract and results report consistent error reduction and stable calibration, yet the manuscript provides insufficient detail on prior specifications for spatial covariance hyperparameters and temporal correlation parameters, as well as on how posterior uncertainty is propagated into the reported RMSE and calibration metrics.
minor comments (2)
  1. [Methods] Notation for the hierarchical model components (e.g., distinction between population-level spatial process and subject-specific deviations) should be clarified with an explicit graphical model or plate diagram to aid reproducibility.
  2. [Simulation study] The simulation scenarios are described at a high level; adding a table that explicitly lists the data-generating parameters for each of the six scenarios would strengthen the claim of robustness.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the detailed and constructive feedback on our manuscript. We address the major comments point by point below, and we plan to incorporate revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Model specification / Results (OASIS-3 application)] The central performance gains (RMSE reductions of 54% and 45%) rest on the assumption that the chosen spatial covariance for the latent deviation process adequately captures neuroanatomical correlation structure in the OASIS-3 parcellation. Without a sensitivity analysis to alternative kernels or graph Laplacians, or a direct comparison of fitted versus empirical spatial correlations, it remains possible that gains arise from added flexibility rather than correct dependence modeling.

    Authors: We thank the referee for highlighting this important point. The spatial covariance function was selected to reflect established patterns of spatial correlation in cortical morphometry, specifically an exponential kernel with range parameter informed by average inter-regional distances in the Desikan-Killiany atlas. Nevertheless, to rule out that the observed improvements are merely due to increased model flexibility, we will perform a sensitivity analysis in the revised manuscript. This will include fitting the model with alternative covariance structures, such as Matérn kernels with varying smoothness parameters and a graph Laplacian based on anatomical adjacency. Additionally, we will add a figure comparing the model-implied spatial correlation matrix with the empirical correlation matrix computed from the OASIS-3 normative sample. These additions will be placed in a new subsection under Results. revision: yes

  2. Referee: [Methods / Simulation study] The abstract and results report consistent error reduction and stable calibration, yet the manuscript provides insufficient detail on prior specifications for spatial covariance hyperparameters and temporal correlation parameters, as well as on how posterior uncertainty is propagated into the reported RMSE and calibration metrics.

    Authors: We agree that the current manuscript lacks sufficient detail on these aspects. In the revised version, we will expand the Methods section to fully specify the prior distributions: for the spatial variance and range parameters we use weakly informative inverse-gamma and half-Cauchy priors, respectively, and for the temporal autocorrelation we employ a uniform prior on [-1,1] for the AR(1) coefficient. Regarding uncertainty propagation, the RMSE values were calculated using the posterior mean of the latent deviation maps as the point estimator, while calibration was assessed via coverage of 95% credible intervals derived from the full posterior. We will clarify this in the text and include additional diagnostics, such as posterior predictive checks for the deviation maps, to demonstrate how uncertainty is accounted for in the performance metrics. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation follows standard Bayesian hierarchical construction

full rationale

The paper defines a unified hierarchical Bayesian model in which the individualized deviation map is introduced as a latent spatial process with an explicit posterior, and the estimator is obtained as the posterior mean under squared-error loss. This is a direct application of standard Bayesian decision theory rather than a reduction of the target quantity to a fitted input or self-referential definition. No equations or claims in the abstract reduce the reported RMSE gains or the spatial-temporal joint modeling to a tautological renaming or self-citation chain; performance is instead demonstrated via external simulation benchmarks and OASIS-3 comparisons. The modeling assumptions (spatial covariance on parcellations, temporal dependence) are conventional and stated as modeling choices, not derived from the results they support.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard Bayesian hierarchical modeling assumptions for spatial processes and longitudinal dependence; no new physical entities are introduced.

free parameters (2)
  • spatial covariance hyperparameters
    Parameters controlling the strength and range of spatial dependence in the latent deviation process; required to define the spatial structure.
  • temporal correlation parameters
    Parameters governing within-subject dependence across repeated visits; fitted or prior-specified within the hierarchical model.
axioms (2)
  • domain assumption The subject-specific deviation map is a latent spatial process with an explicit posterior distribution.
    Invoked when the model treats the individualized deviation map as a latent spatial process yielding a principled Bayes estimator.
  • standard math Squared error loss defines the optimal point estimator for the deviation map.
    Used to justify the posterior mean as the estimator rather than an ad-hoc residual summary.

pith-pipeline@v0.9.0 · 5770 in / 1500 out tokens · 89881 ms · 2026-05-20T21:30:48.169910+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    Bakkour, A., Morris, J. C. & Dickerson, B. C. (2009), ‘The cortical signature of prodromal ad: regional thinning predicts mild ad dementia’,Neurology72(12), 1048–1055

    work page 2009

  2. [2]

    Banerjee, S., Carlin, B. P. & Gelfand, A. E. (2003),Hierarchical modeling and analysis for spatial data, Chapman and Hall/CRC

    work page 2003

  3. [3]

    M., Dinga, R., Kia, S

    Bayer, J. M., Dinga, R., Kia, S. M., Kottaram, A. R., Wolfers, T., Lv, J., Zalesky, A., Schmaal, L. & Marquand, A. (2022), ‘Accommodating site variation in neuroimaging data using normative and hierarchical bayesian models’,NeuroImage264, 119699

    work page 2022

  4. [4]

    (1974), ‘Spatial interaction and the statistical analysis of lattice systems’,Journal of the Royal Statistical Society: Series B (Methodological)36(2), 192–225

    Besag, J. (1974), ‘Spatial interaction and the statistical analysis of lattice systems’,Journal of the Royal Statistical Society: Series B (Methodological)36(2), 192–225

    work page 1974

  5. [5]

    A., Seidlitz, J., White, S

    Bethlehem, R. A., Seidlitz, J., White, S. R., Vogel, J. W., Anderson, K. M., Adamson, C., Adler, S., Alexopoulos, G. S., Anagnostou, E., Areces-Gonzalez, A. et al. (2022), ‘Brain charts for the human lifespan’,Nature604(7906), 525–533. Bucková, B. R., Fraza, C., Rehák, R., Kolenič, M., Beckmann, C. F., Španiel, F., Marquand, A. F. & Hlinka, J. (2025), ‘Us...

    work page 2022

  6. [6]

    & Riddell, A

    Guo, J., Li, P. & Riddell, A. (2017), ‘Stan: A probabilistic programming language’,Journal of statistical software76, 1–32. S19 De Oliveira, V. (2012), ‘Bayesian analysis of conditional autoregressive models’,Annals of the Institute of Statistical Mathematics64(1), 107–133

    work page 2017

  7. [7]

    I., Blacker, D., Rosas, H

    Wright, C. I., Blacker, D., Rosas, H. D. et al. (2009), ‘The cortical signature of alzheimer’s disease: regionally specific cortical thinning relates to symptom severity in very mild to mild ad dementia and is detectable in asymptomatic amyloid-positive individuals’,Cerebral cortex 19(3), 497–510

    work page 2009

  8. [8]

    J., Goldstein, J., Kennedy, D

    Seidman, L. J., Goldstein, J., Kennedy, D. et al. (2004), ‘Automatically parcellating the human cerebral cortex’,Cerebral cortex14(1), 11–22

    work page 2004

  9. [9]

    J., Dinga, R., Beckmann, C

    Fraza, C. J., Dinga, R., Beckmann, C. F. & Marquand, A. F. (2021), ‘Warped bayesian linear regression for normative modelling of big data’,NeuroImage245, 118715

    work page 2021

  10. [10]

    & Marquand, A

    Gaiser, C., Berthet, P., Kia, S., Frens, M., Beckmann, C., Muetzel, R. & Marquand, A. F. (2024), ‘Estimating cortical thickness trajectories in children across different scanners using transfer learning from normative models’,Human Brain Mapping45(2), e26565

    work page 2024

  11. [11]

    (2006), ‘Prior distributions for variance parameters in hierarchical models (comment on article by browne and draper)’, pp

    Gelman, A. (2006), ‘Prior distributions for variance parameters in hierarchical models (comment on article by browne and draper)’, pp. 515–534. Huertas,I.,Oldehinkel,M.,vanOort,E.S.,Garcia-Solis,D.,Mir,P.,Beckmann,C.F.&Marquand, A. F. (2017), ‘A bayesian spatial model for neuroimaging data based on biologically informed basis functions’,NeuroImage161, 134–148

    work page 2006

  12. [12]

    Beckmann, C. F. & Marquand, A. F. (2020), Hierarchical bayesian regression for multi-site S20 normative modeling of neuroimaging data,in‘International Conference on Medical Image Computing and Computer-Assisted Intervention’, Springer, pp. 699–709

    work page 2020

  13. [13]

    Laird, N. M. & Ware, J. H. (1982), ‘Random-effects models for longitudinal data’,Biometrics pp. 963–974

    work page 1982

  14. [14]

    J., Benzinger, T

    LaMontagne, P. J., Benzinger, T. L., Morris, J. C., Keefe, S., Hornbeck, R., Xiong, C., Grant, E., Hassenstab,J.,Moulder,K.,Vlassenko,A.G.etal.(2019),‘Oasis-3: longitudinalneuroimaging, clinical, and cognitive dataset for normal aging and alzheimer disease’,medrxivpp. 2019–12

    work page 2019

  15. [15]

    F., Kia, S

    Marquand, A. F., Kia, S. M., Zabihi, M., Wolfers, T., Buitelaar, J. K. & Beckmann, C. F. (2019),‘Conceptualizingmentaldisordersasdeviationsfromnormativefunctioning’,Molecular psychiatry24(10), 1415–1424

    work page 2019

  16. [16]

    F., Rezek, I., Buitelaar, J

    Marquand, A. F., Rezek, I., Buitelaar, J. & Beckmann, C. F. (2016), ‘Understanding heterogeneity in clinical cohorts using normative models: beyond case-control studies’,Biological psychiatry 80(7), 552–561

    work page 2016

  17. [17]

    F., Koppelmans, V., Jelsone-Swain, L., Kalra, S

    Mejia, A. F., Koppelmans, V., Jelsone-Swain, L., Kalra, S. & Welsh, R. C. (2022), ‘Longitudinal surface-based spatial bayesian glm reveals complex trajectories of motor neurodegeneration in als’,NeuroImage255, 119180

    work page 2022

  18. [18]

    Polson, N. G. & Scott, J. G. (2012), ‘On the half-cauchy prior for a global scale parameter’

    work page 2012

  19. [19]

    & Held, L

    Rue, H. & Held, L. (2005),Gaussian Markov random fields: theory and applications, Chapman and Hall/CRC

    work page 2005

  20. [20]

    F., Sripada, C., Beckmann, C

    Rutherford, S., Barkema, P., Tso, I. F., Sripada, C., Beckmann, C. F., Ruhe, H. G. & Marquand, A. F. (2022), ‘Evidence for embracing normative modeling’,bioRxivpp. 2022–11

    work page 2022

  21. [21]

    Verdi, S., Ruhe, H. G. et al. (2022), ‘The normative modeling framework for computational psychiatry’,Nature protocols17(7), 1711–1734

    work page 2022

  22. [22]

    Dehestani, N., Ng, E. K.-K. et al. (2025), ‘Spectral normative modeling of brain structure’, medRxivpp. 2025–01

    work page 2025

  23. [23]

    M., Yong, K

    Verdi, S., Kia, S. M., Yong, K. X., Tosun, D., Schott, J. M., Marquand, A. F. & Cole, J. H. (2023), ‘Revealingindividualneuroanatomicalheterogeneityinalzheimerdiseaseusingneuroanatomical normative modeling’,Neurology100(24), e2442–e2453

    work page 2023

  24. [24]

    F., Schott, J

    Verdi, S., Marquand, A. F., Schott, J. M. & Cole, J. H. (2021), ‘Beyond the average patient: how neuroimaging models can address heterogeneity in dementia’,Brain144(10), 2946–2953

    work page 2021

  25. [25]

    F., Cole, J

    Marquand, A. F., Cole, J. H. & Initiative, A. D. N. (2024), ‘Personalizing progressive changes to brain structure in alzheimer’s disease using normative modeling’,Alzheimer’s & Dementia 20(10), 6998–7012. S22

    work page 2024