Imputation-free transformer learning enables robust Alzheimer's disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts

Christelle Schneuwly Diaz; Duy-Cat Can; Duy-Thanh Vu; Gilles Allali; Narmina Baghirova; Oliver Y. Ch\'en; Philippe Ryvlin

An imputation-free transformer predicts Alzheimer’s diagnosis and cognitive scores with calibrated uncertainty from incomplete clinical records across cohorts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-07-15 08:43 UTC pith:YSOLMUBI

load-bearing objection Sensible multi-cohort AD setup for a real missing-data problem, but abstract-only so the calibration gains remain unverified claims. the 4 major comments →

arxiv 2607.11656 v2 pith:YSOLMUBI submitted 2026-07-13 q-bio.NC cs.LG

Imputation-free transformer learning enables robust Alzheimer's disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts

Christelle Schneuwly Diaz , Narmina Baghirova , Duy-Thanh Vu , Duy-Cat Can , Gilles Allali , Philippe Ryvlin , Oliver Y. Ch\'en This is my paper

classification q-bio.NC cs.LG

keywords Alzheimer's diseaseimputation-free learningtransformeruncertainty quantificationmultimodal clinical datacohort shiftcalibrationmasked attention

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Incomplete, heterogeneous clinical data block reliable Alzheimer’s diagnosis and severity prediction, and conventional imputation adds bias and overconfidence. This paper introduces NITROGEN, a transformer that never fills missing values; instead it uses masked attention inside each patient and intersample attention across patients so it can learn directly from partially observed multimodal records. Trained on ADNI and tested on independent OASIS-3 and AIBL cohorts, NITROGEN remains competitively discriminative while showing clearer calibration and uncertainty advantages over tree-based ensembles under cohort shift. A modality-aware uncertainty adjustment further raises predictive uncertainty when important modalities are absent. If the claim holds, clinical AI for Alzheimer’s should be judged on calibration, interpretability, and cross-cohort reliability—not accuracy alone.

Core claim

NITROGEN, an imputation-free transformer that jointly models within-patient feature dependencies via masked attention and between-patient structure via intersample attention, enables robust Alzheimer’s diagnostic classification and cognitive-score prediction with calibrated uncertainty across heterogeneous cohorts, without the systematic bias of conventional imputation.

What carries the argument

Masked and intersample attention inside the NITROGEN transformer: masked attention lets the model use whatever features are observed for a given patient, while intersample attention captures relational structure across patients, so missing values never need to be imputed.

Load-bearing premise

Masked within-patient attention plus intersample attention can learn clinically valid structure from partial records without any formal model of why data are missing, and that structure transfers across cohorts whose missingness patterns differ.

What would settle it

On a multi-site held-out cohort with systematically different missingness (for example, imaging or APOE routinely absent), if NITROGEN’s calibration (ECE or reliability diagrams) is no better than a carefully imputed tree ensemble while discrimination also collapses, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Clinical models can train and deploy on incomplete multimodal records without an imputation step that distorts feature relationships.
Predictive confidence can be raised when high-importance modalities are absent, giving clinicians a usable uncertainty signal.
Performance drop on more distributionally different cohorts becomes a measurable reliability check rather than a silent failure.
Cortical thickness in the temporal pole, age, and APOE genotype are consistently informative for AD classification, though none is sufficient alone.
Diagnostic AI evaluation protocols should report calibration and cross-cohort transfer alongside accuracy.

Where Pith is reading between the lines

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

If attention recovers valid structure without a missingness model, similar architectures could apply to other chronic diseases with sparse multi-site health records.
Importance-weighted missingness penalties like the modality-aware adjustment could improve trustworthiness of any multimodal medical predictor.
External validation on community versus research clinics with deliberately mismatched missingness would test whether calibration gains are method-driven or data-pattern-driven.
Pairing the learned feature importances with longitudinal trajectories could yield earlier risk scores that stay calibrated as more modalities appear over time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Sensible multi-cohort AD setup for a real missing-data problem, but abstract-only so the calibration gains remain unverified claims.

read the letter

Punchline: this is a cleanly framed clinical-ML paper on a real deployment barrier—missing multimodal AD data and overconfident models—but we only have the abstract, so the load-bearing claim (imputation-free attention yields genuine calibration under cohort shift) is still uncheckable.

What is actually new is the package, not any single piece. Masked attention for missing features and intersample attention both have prior art; the contribution is applying them jointly as NITROGEN on ADNI (N=7858) with external tests on OASIS-3 and AIBL, plus a modality-aware uncertainty adjustment that scales confidence by the importance of absent modalities. The evaluation story is the right one: train on ADNI, hold out two independent cohorts, and report calibration and uncertainty alongside discrimination rather than accuracy alone. Feature importance (temporal pole thickness, age, APOE) is presented as post-hoc and not individually sufficient, which is honest.

Soft spots are proportional to what we can see. Without ECE/Brier numbers, reliability diagrams, missingness rates, ablations of intersample attention, or a clear account of how “absent” is encoded and how tree ensembles handle missingness, we cannot tell whether the reported calibration edge is real or an artifact of cohort-specific missingness patterns. The abstract asserts robust calibration under shift and expected degradation on more distant cohorts; that is plausible, but free parameters (architecture, uncertainty scaling) and the lack of a formal MAR/MNAR discussion leave the transfer story open. Circularity burden looks low for this empirical setup; the risk is unreproducible detail and baseline under-comparison, not tautology.

Who it is for: people building multimodal AD models or tabular transformers under real-world incompleteness. It deserves a serious referee if the full paper ships the metrics, ablations, and missingness analysis the abstract promises. I would not cite or bring it to reading group on the abstract alone. Send it to peer review rather than desk-reject; the problem and design justify referee time even if revision will be heavy on quantification.

Referee Report

4 major / 3 minor

Summary. The manuscript proposes NITROGEN, an imputation-free transformer that models within-patient feature dependencies and between-patient relational structure via masked and intersample attention, for Alzheimer’s diagnostic classification and cognitive-score prediction from partially observed multimodal records. It is trained on ADNI (N=7858 scans) and evaluated on external cohorts OASIS-3 (N=2675) and AIBL (N=1286). The abstract claims robust calibration and uncertainty-quantification advantages over tree-based ensembles under cohort shift, competitive discrimination, post-hoc identification of temporal-pole thickness, age, and APOE as important features, and a modality-aware uncertainty adjustment that scales predictive uncertainty with the importance of absent modalities.

Significance. If substantiated with quantitative evidence, the work would address a clinically material problem: imputation-induced bias and overconfidence under real-world incompleteness in AD modelling. Emphasizing calibration, cross-cohort reliability, and uncertainty—not accuracy alone—is appropriate for deployment-oriented clinical ML. The multi-cohort external-test design and the explicit modality-aware uncertainty mechanism are, in principle, strengths. Significance hinges on whether the reported calibration gains under shift are real, transferable, and attributable to the proposed attention design rather than evaluation or missingness artifacts.

major comments (4)

The central claim of robust calibration and uncertainty advantages under cohort shift cannot be assessed from the abstract alone: no ECE, Brier score, reliability diagrams, confidence intervals, or statistical comparisons versus tree ensembles are reported. These metrics are load-bearing for the paper’s contribution and must appear with full experimental detail in the manuscript.
The architecture claim rests on masked within-patient plus intersample attention learning clinically valid structure without a formal missingness model (MAR/MNAR). The abstract does not state how absent modalities are marked, what missingness rates or patterns exist across ADNI/OASIS-3/AIBL, or whether gains could be artifacts of cohort-specific incompleteness. A missingness analysis and encoding description are required.
No ablation of intersample (between-patient) attention versus masked-only attention is described. Because intersample attention is an invented, load-bearing component and carries leakage/cohort-artifact risk, its incremental contribution to discrimination and calibration on external cohorts must be quantified.
Baseline fairness is unclear: tree ensembles often handle missingness natively, whereas the abstract contrasts NITROGEN with imputation-based practice. The manuscript must specify exactly how each baseline treats missing values and ensure the comparison isolates the proposed imputation-free design rather than differing missingness handling.

minor comments (3)

Abstract phrasing ‘N=7858 scans’ (and analogous for external cohorts) should be clarified as subjects vs. visits/scans to avoid double-counting concerns in longitudinal AD cohorts.
Feature-importance statements (temporal pole thickness, age, APOE) should be tied to a named method (e.g., attention rollout, permutation, SHAP) and to task (classification vs. score prediction) once full results are presented.
The modality-aware uncertainty adjustment is introduced only at a high level; coefficients, how modality importance is estimated, and whether adjustment is post-hoc or trained should be specified for reproducibility.

Circularity Check

0 steps flagged

No circularity: empirical ML train/test evaluation with no derivation that reduces a claimed prediction to its inputs by construction.

full rationale

This is an abstract-only review of an empirical machine-learning paper. The claimed results are performance metrics (discrimination, calibration, uncertainty) of a proposed architecture (NITROGEN: masked + intersample attention) trained on ADNI and evaluated on held-out external cohorts (OASIS-3, AIBL), with comparisons to tree-based ensembles and a post-hoc feature-importance analysis. There is no mathematical derivation chain in which a 'prediction' or 'first-principles result' is defined in terms of the quantity being predicted, no fitted parameter renamed as an independent prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation. Feature importance (temporal pole thickness, age, APOE) is reported as post-hoc analysis, not as a circular proof of the method. Residual risks (missingness mechanism, evaluation design, calibration metric details) are correctness/generalization concerns, not circularity. Score 0 is the honest finding for a self-contained empirical evaluation against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

Abstract-only audit: free parameters are the usual ML knobs (architecture depth/width, attention design, uncertainty scaling) not numerically disclosed. Domain axioms include standard supervised learning on labeled AD cohorts and the premise that attention over observed tokens plus intersample structure substitutes for explicit imputation. No new physical entity is invented; NITROGEN is an architectural construct. Independent evidence for the method is claimed via external cohorts but not inspectable here.

free parameters (3)

Transformer architecture hyperparameters (depth, heads, embedding size, masking schedule)
Any transformer of this type depends on capacity and training schedule choices that are fitted or tuned on ADNI; values not given in abstract.
Modality-aware uncertainty scaling coefficients
Abstract states uncertainty is augmented proportionally to importance of absent modalities; the proportionality rule/weights are free design choices that affect calibration claims.
Feature/modality importance weights used in uncertainty adjustment
Importance of absent modalities must be estimated or assigned; abstract does not specify whether these are learned, fixed from ADNI, or hand-set.

axioms (4)

domain assumption Partially observed multimodal clinical records can be modeled end-to-end with masked attention without an explicit probabilistic missingness model.
Core methodological premise of the imputation-free approach stated in the abstract.
ad hoc to paper Between-patient (intersample) attention provides useful relational signal for individual diagnosis without harmful leakage or cohort artifacts.
Intersample attention is a design choice specific to this method; validity under clinical deployment constraints is assumed.
domain assumption ADNI labels and features are a suitable training distribution for evaluating generalization to OASIS-3 and AIBL.
Standard multi-cohort transfer assumption underlying the external validation design.
domain assumption Tree-based ensemble methods with conventional imputation are the relevant competitive baseline for discrimination and calibration.
Comparison framing in the abstract; other deep tabular or imputation-aware baselines are not mentioned.

invented entities (2)

NITROGEN (imputation-free transformer with masked and intersample attention) no independent evidence
purpose: Jointly model within-patient feature dependencies and between-patient structure from incomplete multimodal AD records.
Named model introduced by the paper; architectural construct rather than a new biological entity. Independent evidence would be released code and full multi-cohort metrics, not available in the abstract.
Modality-aware uncertainty adjustment no independent evidence
purpose: Inflate predictive uncertainty when important modalities are missing to improve calibration under incomplete inputs.
New procedural component claimed to yield calibrated confidence; falsifiable only via reported calibration curves/metrics not present in the abstract.

reviewed 2026-07-15 · how reviews work

0 comments

read the original abstract

Accurate diagnostic classification and disease-severity prediction for Alzheimer's disease are hampered by the incompleteness and heterogeneity of real-world clinical data. Left unaddressed, these barriers prevent reliable disease modelling and hinder effective clinical evaluation. Conventional imputation strategies introduce systematic bias, distort inter-feature relationships, and yield overconfident predictions, limitations especially consequential in diagnostic settings. Here, we propose NITROGEN, an imputation-free transformer that jointly models within-patient feature dependencies and between-patient relational structure through masked and intersample attention, enabling robust multimodal learning directly from partially observed records. We trained NITROGEN on ADNI (N=7858 scans), and evaluated it on two independent cohorts: OASIS-3 (N=2675 scans) and AIBL (N=1286 scans). Across cohorts and diagnostic and cognitive score prediction tasks, NITROGEN showed robust calibration and uncertainty quantification advantages over tree-based ensemble methods, while maintaining competitive discriminative performance. Cross-cohort and cross-method analyses identified cortical thickness in the temporal pole, age, and APOE genotype as important, though not individually sufficient, features for AD classification. We further introduced a modality-aware uncertainty adjustment that augments predictive uncertainty proportionally to the importance of absent modalities, enabling calibrated confidence when diagnostic information is unavailable. Together, our results show that imputation-free attention learning preserved meaningful discrimination under cohort shift, revealing expected degradation on more distributionally different cohorts, and demonstrate that evaluating models along calibration, interpretability, and cross-cohort reliability, not accuracy alone, is essential for clinical deployment.

Review history (2 revisions) →

discussion (0)

This paper was first reviewed by grok-4.5 on July 15, 2026.