pith. sign in

arxiv: 2505.24259 · v2 · submitted 2025-05-30 · 📊 stat.ME

Partially-shared Imaging Regression on Integrating Heterogeneous Brain-Cognition Associations across Alzheimer's Diagnoses

Yang Sui , Qi Xu , Ting Li , Yang Bai , Annie Qu This is my paper

Pith reviewed 2026-05-19 13:40 UTC · model grok-4.3

classification 📊 stat.ME
keywords Alzheimer's diseaseimaging regressionheterogeneous associationspartial sharingbrain-cognition pathwaysADNI dataspatial smoothnessinterpretability
0
0 comments X

The pith

The PAIR model represents imaging coefficients as weighted combinations of smooth spatial components to capture heterogeneous brain-cognition associations across Alzheimer's diagnostic groups.

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

The authors develop a partially-shared imaging regression model for integrating demographic, imaging, and cognitive data from different diagnostic groups in Alzheimer's studies. By expressing imaging coefficients through smooth spatial components and applying penalties for smoothness and selective sharing, the model achieves prediction accuracy on par with deep learning while revealing interpretable differences in brain pathways. In the ADNI application, it finds that hippocampal regions contribute little to cognition among cognitively normal individuals but play a substantial role among the cognitively impaired, especially in specific subfields.

Core claim

PAIR models imaging coefficients as weighted combinations of smooth spatial components. Total variation penalty enforces spatial smoothness and selective integration penalty learns partial-sharing structures across groups. This produces minimax-optimal error bounds that adapt to the underlying sharing pattern and, on ADNI data, uncovers substantial heterogeneity with minimal hippocampal contribution in the cognitively normal group but substantial contribution in the cognitively impaired group, particularly in CA1, CA3, and presubiculum.

What carries the argument

The selective integration penalty, which adaptively learns partial-sharing structures of smooth spatial components for imaging coefficients across diagnostic groups.

If this is right

  • The model attains predictive accuracy comparable to advanced deep learning approaches on ADNI data.
  • It provides superior interpretability by identifying group-specific brain-cognition pathways.
  • Theoretical error bounds dynamically adapt to different degrees of sharing between groups.
  • Application to ADNI data shows hippocampal imaging contributes minimally in cognitively normal subjects but substantially in cognitively impaired subjects.
  • Particular subfields like CA1, CA3, and presubiculum are highlighted as important in the impaired group.

Where Pith is reading between the lines

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

  • This framework could be applied to other diseases with heterogeneous imaging associations to discover group-specific biomarkers.
  • Longitudinal extensions might test whether the learned partial-sharing patterns predict future cognitive decline.
  • The emphasis on spatial smoothness and partial sharing suggests similar benefits for other high-dimensional imaging modalities beyond MRI.

Load-bearing premise

The imaging coefficients are well-approximated by linear combinations of a few smooth spatial components, and the selective integration penalty accurately recovers the true partial-sharing structure between groups.

What would settle it

Fitting the model to a simulated dataset where the true coefficients do not follow the smooth component structure or where groups are fully separate would show if the error bounds and heterogeneity findings hold or break.

read the original abstract

Alzheimer's Disease Neuroimaging Initiative (ADNI) diagnostic groups present strong heterogeneous associations among demographic, imaging, and cognitive data. We propose a novel PArtially-shared Imaging Regression (PAIR) model to represent imaging coefficients as weighted combinations of smooth spatial components. A Total Variation penalty is applied to enforce spatial smoothness, and a Selective Integration penalty is introduced to adaptively learn partial-sharing structures across groups. Theoretically, we establish minimax-optimal error bounds that dynamically adapt to varying sharing paradigms. Numerically, PAIR achieves predictive accuracy comparable to advanced deep learning models while providing superior interpretability. Applied to ADNI data, PAIR reveals substantial heterogeneity in brain-cognition pathways between cognitively normal (CN) and cognitively impaired (CI) groups, with hippocampal imaging contributing minimally in the CN group but substantially in the CI group, particularly in the CA1, CA3, and presubiculum subfields.

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

3 major / 2 minor

Summary. The paper proposes the PArtially-shared Imaging Regression (PAIR) model to integrate heterogeneous brain-cognition associations across Alzheimer's diagnostic groups in the ADNI dataset. Imaging coefficients are represented as weighted combinations of smooth spatial components, with a Total Variation penalty enforcing spatial smoothness and a Selective Integration penalty adaptively learning partial-sharing structures across groups. The authors claim to establish minimax-optimal error bounds that dynamically adapt to varying sharing paradigms, achieve predictive accuracy comparable to advanced deep learning models with superior interpretability, and apply the model to ADNI data to reveal substantial heterogeneity, with hippocampal imaging contributing minimally in the cognitively normal (CN) group but substantially in the cognitively impaired (CI) group, particularly in the CA1, CA3, and presubiculum subfields.

Significance. If the theoretical bounds are rigorously established with explicit derivations and the selective integration penalty is shown via controlled simulations to correctly recover known partial-sharing structures, the work would offer a valuable interpretable framework for multi-group neuroimaging regression with adaptive theoretical guarantees. This could advance statistical methods for heterogeneous data in Alzheimer's research by balancing predictive performance with mechanistic insights into brain-cognition pathways, though current gaps in validation and detail reporting limit immediate impact.

major comments (3)
  1. Theoretical analysis section: The manuscript claims minimax-optimal error bounds that dynamically adapt to varying sharing paradigms, yet provides no derivation details, explicit assumptions, or proof sketches for how the bounds adapt to the selective integration penalty and partial-sharing structure. This is load-bearing for the central theoretical contribution and prevents assessment of whether the bounds hold independently of the data-driven penalty tuning.
  2. Numerical experiments and ADNI application sections: No simulation experiments with known ground-truth sharing patterns are reported to validate that the selective integration penalty recovers partial-sharing structures without inducing spurious group differences. This directly undermines the reliability of the reported heterogeneity findings (minimal hippocampal contribution in CN vs. substantial in CI, especially CA1/CA3/presubiculum), as the central modeling premise that coefficients are weighted combinations of smooth components and the penalty learns true partial-sharing lacks controlled testing.
  3. Data and methods section: The manuscript does not specify data exclusion rules, preprocessing details, or cross-validation procedures for the ADNI dataset, and the predictive accuracy comparisons to deep learning models lack error bars, standard deviations, or statistical tests, making it difficult to evaluate the claimed comparability.
minor comments (2)
  1. Abstract: The claim of 'numerical comparability to advanced deep learning models' should specify the exact models, metrics (e.g., MSE, R²), and datasets used for comparison.
  2. Notation and presentation: The definition of the selective integration penalty and its relation to the weighted combination of spatial components could be clarified with an explicit equation or diagram early in the model section.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us identify areas for improvement. We address each major comment below and commit to revisions that strengthen the clarity and validation of the work without altering its core contributions.

read point-by-point responses

  1. Referee: Theoretical analysis section: The manuscript claims minimax-optimal error bounds that dynamically adapt to varying sharing paradigms, yet provides no derivation details, explicit assumptions, or proof sketches for how the bounds adapt to the selective integration penalty and partial-sharing structure. This is load-bearing for the central theoretical contribution and prevents assessment of whether the bounds hold independently of the data-driven penalty tuning.

    Authors: We agree that the current presentation of the theoretical results is high-level and lacks sufficient detail for independent verification. In the revised manuscript, we will expand the theoretical analysis section to include the full set of assumptions, a detailed derivation outline, and proof sketches demonstrating how the minimax bounds adapt to the selective integration penalty and different partial-sharing regimes. These additions will explicitly address independence from penalty tuning. revision: yes

  2. Referee: Numerical experiments and ADNI application sections: No simulation experiments with known ground-truth sharing patterns are reported to validate that the selective integration penalty recovers partial-sharing structures without inducing spurious group differences. This directly undermines the reliability of the reported heterogeneity findings (minimal hippocampal contribution in CN vs. substantial in CI, especially CA1/CA3/presubiculum), as the central modeling premise that coefficients are weighted combinations of smooth components and the penalty learns true partial-sharing lacks controlled testing.

    Authors: We concur that controlled simulations with known ground-truth structures are necessary to validate the selective integration penalty. The revised version will incorporate a new simulation study that generates data under predefined partial-sharing patterns. We will show that the penalty recovers the true structures with high accuracy and does not induce spurious differences, thereby bolstering confidence in the ADNI heterogeneity results. revision: yes

  3. Referee: Data and methods section: The manuscript does not specify data exclusion rules, preprocessing details, or cross-validation procedures for the ADNI dataset, and the predictive accuracy comparisons to deep learning models lack error bars, standard deviations, or statistical tests, making it difficult to evaluate the claimed comparability.

    Authors: We will substantially expand the Data and Methods section to document the ADNI data exclusion criteria, all preprocessing steps for imaging and cognitive variables, and the cross-validation protocol. In addition, the predictive accuracy tables will be updated to include standard deviations across folds and results from statistical tests comparing PAIR to the deep learning baselines. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines the PAIR model by expressing imaging coefficients as weighted combinations of smooth spatial components, introduces a Total Variation penalty for spatial smoothness and a Selective Integration penalty to learn partial-sharing structures, derives minimax-optimal error bounds that adapt to sharing paradigms, and applies the fitted model to ADNI data to report CN/CI heterogeneity findings. This chain is self-contained: the model and penalties are proposed as a novel construction, the theoretical bounds follow from the stated assumptions without reducing to data-specific fits by construction, and the reported pathway differences are empirical outputs rather than presupposed inputs. No load-bearing step equates a prediction or bound to its own fitted parameters, and the provided text contains no self-citations invoked for uniqueness theorems or ansatzes. The selective integration penalty is a deliberate modeling choice whose behavior is analyzed theoretically and tested empirically, not a tautological redefinition.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The model rests on domain assumptions about spatial smoothness and partial sharing that are not independently verified outside the ADNI application; no free parameters are explicitly listed but tuning of penalties is implied.

free parameters (1)
  • penalty tuning parameters
    Total variation and selective integration penalties require hyperparameters chosen or fitted to achieve the reported smoothness and sharing structure.
axioms (2)
  • domain assumption Imaging coefficients admit representation as weighted sums of smooth spatial components
    Stated as the core modeling step in the abstract.
  • domain assumption Selective integration penalty can adaptively recover partial-sharing patterns across diagnostic groups
    Required for the heterogeneity analysis and theoretical adaptation claim.

pith-pipeline@v0.9.0 · 5688 in / 1399 out tokens · 53846 ms · 2026-05-19T13:40:21.807343+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.