pith. sign in

arxiv: 2604.07764 · v1 · submitted 2026-04-09 · 📊 stat.ME

Bayesian Tensor-on-Tensor Varying Coefficient Model for Forecasting Alzheimer's Disease Progression

Yajie Liu , Hengrui Luo , Suprateek Kundu This is my paper

Pith reviewed 2026-05-10 18:19 UTC · model grok-4.3

classification 📊 stat.ME
keywords Bayesian tensor modelingAlzheimer's disease progressionlongitudinal MRIGaussian processescortical thickness predictionspatial heterogeneityvarying coefficient modelbrain aging forecasting
0
0 comments X

The pith

A Bayesian tensor-on-tensor model forecasts future cortical thickness and brain aging in Alzheimer's from longitudinal MRI.

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 framework to predict how brain structures evolve over time in Alzheimer's patients by treating image data as tensors. It incorporates Gaussian process priors to model nonlinear voxel relationships, low-rank tensor coefficients to enforce spatial structure, and patch-to-voxel mappings to handle local heterogeneity. This produces forecasted future brain images that can be used to track aging processes. A reader would care because accurate forecasts could support earlier detection of progression and better understanding of disease mechanisms in real patient data. Simulations confirm advantages in estimation and prediction, and the model is demonstrated on T1-weighted MRIs from the Alzheimer's Disease Neuroimaging Initiative.

Core claim

The authors claim that a Bayesian tensor-on-tensor varying coefficient model with low-rank tensor coefficients and Gaussian process priors flexibly captures nonlinear voxel-level relationships and spatial heterogeneity through patch-to-voxel mappings. An efficient parallel MCMC algorithm is provided for posterior sampling. When applied to longitudinal T1-weighted MRIs from the ADNI study, the resulting predictions accurately forecast future cortical thickness and support reliable prediction of brain aging, indicating biological relevance for tracking neurobiological changes.

What carries the argument

Bayesian tensor-on-tensor varying coefficient model that uses low-rank tensor coefficients for spatial structure, Gaussian process priors for nonlinearity, and patch-to-voxel mappings for spatial heterogeneity.

If this is right

  • The approach yields more accurate coefficient estimation and statistical inference than existing methods for high-dimensional image data.
  • Simulations show improved prediction accuracy and better scalability to large images.
  • Applied to ADNI data the model produces accurate forecasts of future cortical thickness.
  • The predicted images enable reliable assessment of brain aging with clear biological relevance.

Where Pith is reading between the lines

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

  • The same structure could be tested on imaging data from other neurodegenerative conditions to forecast progression.
  • Linking the forecasted images to genetic or fluid biomarker data might improve individualized risk estimates.
  • The parallel MCMC design suggests the method could handle even larger multi-modal datasets in future work.
  • Independent validation on separate patient cohorts would test whether the forecasts generalize beyond the ADNI sample.

Load-bearing premise

The model assumes low-rank tensor coefficients together with patch-to-voxel mappings and Gaussian process priors are sufficient to represent the nonlinear voxel relationships and spatial patterns in brain imaging data without substantial bias or overfitting.

What would settle it

A new set of longitudinal MRI scans from Alzheimer's patients where the model's predicted cortical thickness maps show large systematic differences from the actual follow-up images or where derived brain aging predictions fail to correlate with clinical cognitive decline measures.

Figures

Figures reproduced from arXiv: 2604.07764 by Hengrui Luo, Suprateek Kundu, Yajie Liu.

Figure 1
Figure 1. Figure 1: Illustration of the BTOT-VC model structure and 3D voxel-centered patch construction. (a) A schematic representation of the BTOT-VC model (shown without covariates for clarity). Each cube denotes a 3D tensor, with unit cubes corresponding to voxels. The model is expressed as Yn = Γ + Θ ⊙ Mn,·(XP,n) + En, n = 1, . . . , N, highlighting how voxel-wise dependencies between input and output tensors are capture… view at source ↗
Figure 2
Figure 2. Figure 2: ROI-Level Longitudinal Prediction Accuracy. Panels (a)– (b) summarize voxel-wise prediction accuracy for cortical thickness across 83 ROIs, using correlation (a), RPE (b). For each ROI, boxplots display the distribution of correlation and RPE values across voxels. Cortical thickness values were extracted using the DKT atlas, and all models were trained and evaluated independently within each group and ROI.… view at source ↗
Figure 3
Figure 3. Figure 3: DIC values for different tensor rank choices for simulation setting 3.a.ii [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example traceplots for parameters in simulation. The blue curve corresponds to [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
read the original abstract

We propose a novel tensor-on-tensor modeling framework that flexibly models nonlinear voxel-level relationships using Gaussian process (GP) priors, while incorporating the spatial structure of the output tensor through low-rank tensor-based coefficients. Spatial heterogeneity is captured through patch-to-voxel mappings, enabling each output voxel to depend on its spatial neighborhood. The proposed interpretable and flexible Bayesian tensor-on-tensor framework is able to capture nonlinearity, spatial information, and spatial heterogeneity. We develop an efficient Markov chain Monte Carlo (MCMC) algorithm that exploits parallel structure to sample voxel-specific GP atoms and update low-rank tensor coefficients. Extensive simulations reveal advantages of the proposed approach over existing methods in terms of coefficient estimation, inference, prediction, and scalability to high-dimensional images. Applied to longitudinal image prediction with T1-weighted MRIs from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the proposed method can accurately forecast future cortical thickness. The predicted images also enable reliable prediction of brain aging, underscoring their biological relevance. Overall, the ADNI analysis highlights the model's ability to forecast future neurobiological changes that has important implications for early detection of AD.

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 / 3 minor

Summary. The paper proposes a Bayesian tensor-on-tensor varying coefficient model for longitudinal neuroimaging that uses Gaussian process priors to capture nonlinear voxel-level relationships, low-rank tensor coefficients to encode spatial structure in the output tensor, and patch-to-voxel mappings to handle spatial heterogeneity. An efficient MCMC algorithm is developed that exploits parallel sampling of voxel-specific GP atoms and updates to the low-rank coefficients. Simulations demonstrate advantages over existing methods in coefficient estimation, inference, prediction accuracy, and scalability to high-dimensional images. On T1-weighted MRI data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the model is shown to accurately forecast future cortical thickness, with the resulting predictions enabling reliable quantification of brain aging that has biological relevance for early Alzheimer's detection.

Significance. If the central claims hold, the work advances statistical methodology for tensor-valued longitudinal data by providing a flexible, interpretable Bayesian framework that jointly addresses nonlinearity, spatial dependence, and heterogeneity—challenges that are central to neuroimaging applications. The efficient MCMC implementation that leverages parallel structure, together with the simulation studies and the ADNI application demonstrating forecasting of cortical thickness and brain aging, constitute concrete strengths. These elements position the approach as potentially useful for early detection and monitoring of Alzheimer's progression, with broader implications for tensor regression in medical imaging.

major comments (2)
  1. [ADNI Analysis section] ADNI Analysis section: the claim that the method 'can accurately forecast future cortical thickness' and yields 'reliable prediction of brain aging' is load-bearing for the headline result, yet the manuscript provides no quantitative uncertainty measures (e.g., posterior credible intervals derived from the GP priors) or external validation against clinical AD biomarkers to substantiate accuracy beyond point predictions.
  2. [Methods, patch-to-voxel mapping definition] Methods, patch-to-voxel mapping definition: the assertion that patch-to-voxel mappings capture spatial heterogeneity without introducing bias rests on an untested modeling choice; no sensitivity analysis to patch size or neighborhood definition is reported, which directly affects the central claim that the framework flexibly handles real brain imaging spatial structure.
minor comments (3)
  1. [MCMC Algorithm section] The abstract states that the MCMC 'exploits parallel structure' but the main text does not detail the specific parallelization strategy or software implementation, which would aid reproducibility.
  2. [Simulation studies section] Simulation studies section: the competing methods and image dimensions used in the comparisons are not enumerated in the abstract or early summary; adding these details would clarify the scope of the reported advantages.
  3. Notation for the low-rank tensor coefficients and GP atoms should be introduced with a single consolidated table or diagram to reduce cross-referencing across sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments and positive overall assessment. We address each major comment below and outline the revisions we will incorporate to strengthen the manuscript.

read point-by-point responses

  1. Referee: [ADNI Analysis section] ADNI Analysis section: the claim that the method 'can accurately forecast future cortical thickness' and yields 'reliable prediction of brain aging' is load-bearing for the headline result, yet the manuscript provides no quantitative uncertainty measures (e.g., posterior credible intervals derived from the GP priors) or external validation against clinical AD biomarkers to substantiate accuracy beyond point predictions.

    Authors: We agree that uncertainty quantification strengthens the claims. Because the model is fully Bayesian, posterior samples from the MCMC are already available and credible intervals for predicted cortical thickness can be computed directly from the GP posterior. In the revised manuscript we will add these intervals (both numerical summaries and visualizations for representative regions) to the ADNI results. For external validation, the primary evidence is out-of-sample forecasting against held-out future scans within the ADNI cohort; we will additionally report correlations between the predicted brain-aging trajectories and available clinical measures (e.g., MMSE, CDR) already present in the ADNI database to provide further substantiation. revision: partial

  2. Referee: [Methods, patch-to-voxel mapping definition] Methods, patch-to-voxel mapping definition: the assertion that patch-to-voxel mappings capture spatial heterogeneity without introducing bias rests on an untested modeling choice; no sensitivity analysis to patch size or neighborhood definition is reported, which directly affects the central claim that the framework flexibly handles real brain imaging spatial structure.

    Authors: The patch-to-voxel construction is motivated by the local spatial smoothness typical in neuroimaging, but we acknowledge that its robustness should be demonstrated. In the revised manuscript we will add a sensitivity analysis that varies patch size (e.g., 3×3×3 versus 5×5×5) and neighborhood definitions on both the simulation settings and the ADNI data, reporting the resulting changes in prediction error, coefficient recovery, and computational cost. This will directly support the claim that the framework flexibly accommodates spatial heterogeneity. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The manuscript introduces a new Bayesian tensor-on-tensor varying coefficient model that combines low-rank tensor coefficients, patch-to-voxel mappings, and Gaussian process priors. All load-bearing steps (model specification, MCMC sampling, simulation comparisons, and ADNI forecasting) are presented as independent constructions with external validation via simulations and real-data metrics. No equation reduces to a fitted input by definition, no uniqueness theorem is imported from self-citations, and no ansatz or renaming is smuggled in. The framework remains self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The abstract implies several modeling choices whose details are not specified, including the choice of tensor rank, GP kernel form, and the exact patch size for spatial mapping; these function as free parameters or domain assumptions whose impact on the central forecasting claim cannot be evaluated from the abstract alone.

free parameters (2)
  • tensor rank
    Low-rank tensor coefficients require a chosen rank that controls model complexity and is not specified.
  • GP kernel hyperparameters
    Gaussian process priors for voxel-level nonlinearity require kernel parameters that must be set or estimated.
axioms (2)
  • domain assumption Gaussian processes provide flexible nonparametric modeling of nonlinear functions
    Invoked to capture voxel-level relationships.
  • domain assumption Low-rank tensor decomposition adequately represents spatial structure in brain images
    Used to incorporate spatial information via coefficients.

pith-pipeline@v0.9.0 · 5499 in / 1364 out tokens · 62310 ms · 2026-05-10T18:19:32.797469+00:00 · methodology

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

2 extracted references · 2 canonical work pages

  1. [1]

    Marco Lorenzi, Maurizio Filippone, Giovanni B Frisoni, Daniel C Alexander, Sébastien Ourselin, Alzheimer’s Disease Neuroimaging Initiative, et al

    ISSN 1537-2715. Marco Lorenzi, Maurizio Filippone, Giovanni B Frisoni, Daniel C Alexander, Sébastien Ourselin, Alzheimer’s Disease Neuroimaging Initiative, et al. Probabilistic disease progression modeling to characterize diagnostic uncertainty: application to staging and prediction in alzheimer’s disease.NeuroImage, 190:56–68, 2019. Luise Christine Löwe,...

    work page 2019

  2. [2]

    Then the sampling steps for the full MCMC algorithm are as follows, where ranksr∈[1, R]and dimensiond∈[1, D]are looped through. Step 1: Let Y γ n,r =Y n − ˆΓr −bΘ⊙ \Mn,·(XP,n)− PS s=1cDszns be the rank r specific residual corresponding to the Γ term, where ˆΓr = PR r′=1,r′̸=r ˆγ1·,r′ ◦ · · · ◦ ˆγD·,r′ where ˆγ1·,r′,· · ·, ˆγD·,r′ are sampled from the most...

    work page 2026