NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting
Tobias Susetzky, Raphael Rehms, Dmitrii Seletkov, Özgün Turgut, Michelle Espranita Liman, Lisa Steinhelfer, Rickmer Braren, Daniel Rueckert
cs.LG · cs.AI
Quick verdict
Claims first holistic generative patient-journey model on MIMIC (559M events); broad evaluation, but severe miscalibration and no trained baselines temper it.
Summary
NOAH is a 150M-parameter transformer decoder trained autoregressively on 559 million timestamped clinical events from 299,712 patients and 431,000 hospital visits across the MIMIC family, covering text, chest X-rays, echocardiograms, ECG waveforms, numeric, and categorical data via frozen external encoders. Two architectural novelties are proposed: a bidirectional temporal attention that adds backward time gaps to attention keys and forward gaps to queries, and a variational latent space with a context-conditional prior that routes inference-time stochasticity through a single shared variable, structurally guarded against posterior collapse. Evaluations span generative forecasting with optional time control (macro event-type AUROC 0.829 free, 0.856 time-controlled at 24h versus 0.610 persistence), Monte Carlo zero-shot classification (72-h mortality AUROC 0.972; prolonged stay 0.741; 30-day readmission 0.609), counterfactual fluid simulation whose treatment-effect sign matches the SMART sepsis subgroup but at roughly twice its magnitude, linear probes for 15 ICD chapters and 29 comorbidities (AUROC 0.72-0.94), NEWS2 recovery (MAE 1.336), and DeepSurv survival analysis (C-index up to 0.849). Main limitations: predicted absolute probabilities are badly miscalibrated (simulated mortality 73.7% vs 21.0% observed; ECE 27.7pp), no head-to-head comparison with related foundation models such as APOLLO appears in the paper, single-dataset evaluation, and code not yet released.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 3
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- NOAH is the first truly holistic generative model representing and forecasting the full multimodal patient journey across all clinical modalities and entire histories.
- Bidirectional temporal attention weights history by both recency and prediction horizon, with time control improving forecasting by up to +0.19 AUROC at short horizons.
- A context-conditional variational prior captures trajectory stochasticity, and its KL surprise correlates with clinically meaningful events such as ED arrivals and acute deterioration.
- NOAH performs training-free zero-shot classification of clinical outcomes via Monte Carlo rollouts, reaching AUROC 0.972 for 72-hour mortality.
- Counterfactual fluid simulation reproduces the direction of the SMART trial's sepsis-subgroup treatment effect, at roughly twice its magnitude.
Strengths
Genuinely novel, well-motivated time integration
Bidirectional temporal attention adds backward inter-event gaps to attention keys and forward gaps to queries; the paper argues (correctly) that the forward delta is unavailable from history under causal masking, and the Fig. 6b ablation shows attention shifts most when content-content and query-side temporal terms are removed (TV distance up to ~0.50).
Carefully hardened variational latent space
A learned context-conditional diagonal-Gaussian prior with σ floor 0.05 and squash cap 2, linear KL annealing, and an auxiliary prior-mode reconstruction loss (weight γ) structurally block both degenerate solutions (posterior pass-through, prior variance inflation); the resulting KL surprise tracks clinical novelty, peaking 2.2× at ED arrival (Fig. 2h).
Unusually broad, honestly reported evaluation
Transparency and mechanism analyses
Candid limitations section
Weaknesses
Gross absolute miscalibration of generated outcomes
No trained comparison model
Modality breadth overstated by coverage
Selection effects in Monte Carlo scoring
Reproducibility pending; confounding unaddressed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
NOAH (arXiv 2609.09140; TUM, UKE Hamburg, Imperial) positions itself as the first truly holistic generative model of the patient journey: a task-agnostic, time-aware transformer decoder (150.3M parameters, Suppl. Table 16) trained on 559 million timestamped clinical events from 299,712 patients across 431,000 hospital visits spanning six MIMIC datasets, covering free text, chest X-rays, echocardiograms, ECG waveforms, and numeric/categorical data. Three coupled claims are advanced: (i) bidirectional temporal attention — backward gaps added to attention keys, forward gaps to queries — makes attention sensitive to the prediction horizon; (ii) a variational latent space with a context-conditional prior supplies controlled inference-time stochasticity; (iii) one pretrained model serves generative forecasting with time control, zero-shot Monte Carlo outcome classification, counterfactual intervention simulation, and representation probing.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is NOAH and what data was it trained on?
NOAH is a 150M-parameter generative transformer that models a patient's entire multimodal medical history as a sequence of timestamped events. It was trained on 559 million clinical events from 431,000 hospital visits of about 299,000 patients using the MIMIC dataset family (structured records, notes, chest X-rays, echocardiograms, ECGs), and supports forecasting future events, patient-state embeddings, and intervention simulation.
Can NOAH predict patient outcomes without task-specific training?
Partially. Via Monte Carlo rollouts it reaches zero-shot AUROC 0.972 for 72-hour mortality and 0.741 for prolonged stay, but 30-day readmission is weak (0.609), predicted probabilities are poorly calibrated (mortality overpredicted roughly 3.5x in the counterfactual setting), and 18-35% of simulated rollouts are discarded as invalid.
Does NOAH simulate treatment effects reliably?
Its counterfactual simulation of saline vs lactated Ringer's in sepsis matches the direction of the SMART trial's effect but at roughly twice the magnitude, and it overpredicts absolute mortality (73.7% predicted vs 21.0% observed). The authors position it for hypothesis generation, not clinical decisions, and explicitly state it is not causal modeling.
Is NOAH's code or data publicly available?
The data are the credentialed-access MIMIC datasets distributed via PhysioNet and cannot be redistributed. The paper states code will be released on GitHub for academic research upon publication; as of this preprint (arXiv 2609.09140, September 2026) no repository is live.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with four marked sampling gaps: part of the Results narrative after Figure 3, the start of Methods (dataset preprocessing and text/CXR encoder details), part of the zero-shot/counterfactual evaluation configuration, and portions of Supplementary Tables 9-10; judgments touching those sections rest on surrounding context.
- Figure axis values (Figures 2-6) are only partially recoverable from extracted text; quantitative readings combine figure captions, embedded numbers, and the reproduced tables.
- The arXiv stamp (2609.09140, 8 Sep 2026) indicates a very recent preprint; peer-review status cannot be determined from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.