acceptodds
Under review as a conference paper at ICLR 2027

PulmoWorld: An Asynchronous Multimodal Medical World Model for Longitudinal Pneumonia Dynamics

Abstract

Pneumonia changes across respiration, inflammation, and organ function, while the measurements that describe it arrive at different times and are often incomplete. We introduce PULMOWORLD, a world model for longitudinal pneumonia dynamics that connects asynchronous state estimation, recursive simulation, and evidence-constrained planning. History-aware filtering separates state propagation from observation-driven correction, so clinical, imaging, and bronchoalveolar lavage evidence update a shared belief only when observed. That belief supports recursive forecasts of clinical values and modality representations, and summaries of future prognosis. Cap-MBRL searches imagined continuations inside an authorized action set and replaces a frozen fallback only when a candidate meets the risk budget and a predicted-cost margin. On the pneumonia cohort MASS, recursive forecasts are competitive with task-trained sequence models, and simulated future states retain prognostic information. Trained separately on public MIMIC-IV-CXR chest radiographs and clinical trajectories, this model improves prognostic discrimination over latest-value fusion in development and is competitive on long-horizon error on the locked pneumonia panel. In five known-mechanism MASS-Syn environments, authorized search lowers selection regret relative to direct fallback execution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
world model;Pulmonary Evolution; longitudinal patient modeling, clinical time series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_Wa9NOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces PULMOWORLD, which incorporates synchronous multimodal longitudinal data to estimate a patient's state, simulate future states, and ideally perform evidence-constrained planning. PULMOWORLD builds a Gaussian belief of a patient's state using past observations and evolves it with new observation corrections. Validated on a private MASS dataset and the public MIMIC dataset, PULMOWORLD's simulated states retain clinical and prognostic information. For evidence-constrained planning, the paper introduces Cap-MBRL, which restricts policy optimization to actions passing predictive-evidence and identification-evidence checks, and replaces a frozen fallback policy only when a candidate satisfies a risk budget and predicted-cost margin. On the synthetic dataset MASS-Syn with known mechanism environments, Cap-MBRL lowers selection regret relative to direct fallback execution.

Critical strengths
  1. Handling asynchronous and incomplete multimodal Data: PULMOWORLD decoupled elapsed-time state propagation from observation corrections and xplicitly modeling missing modalities with masks, allowing flexible integration of irregular, asynchronous longitudinal clinical measurements (such as imaging and BALF).
  2. Multi-faceted evaluation across core world-model capabilities. PULMOWORLD separately examines state modeling through history reconstruction, future simulation through recursive clinical and modality-representation forecasting and prognostic utility, and planning in known-mechanism environments.
Critical weaknesses
  1. The empirical evidence for nontrivial planning performance is restricted to the synthetic MASS-Syn environments. On the real cohorts, planning remains fail-closed because the identification gate is not passed. The clinical relevance and generalizability of the reported planning results therefore remain unclear.
  2. The construction of MASS-Syn and its relationship to the real MASS cohort and pneumonia physiology are not sufficiently specified or validated.
Questions for authors
  1. Why is each latent state represented by a diagonal Gaussian? Is there any theoretical or empirical justification? Can the authors compare it with deterministic, non-Gaussian, or point-process alternatives?
  2. How does removing CT, CXR, or BALF affect the latent state, future predictions, and action selection? Can the authors report clinical-only, clinical-plus-imaging, and full-modality results, as well as an ablation removing cross-compartment messages?
  3. How were Qwen3-VL-30B-A3B and Hulu-Med-32B adapted for longitudinal numerical forecasting? Please clarify the input/output format, prompting or fine-tuning procedure, and why their errors are unusually large. For these models, H4 error is lower than H1; how should this pattern be interpreted?
  4. Please report hardware, memory, training time, preprocessing cost, and inference overhead. Also, please explain the rationale for using the selected pretrained encoders and freezing them. Why were stronger CXR-specific pretrained encoders or task-specific fine-tuning not considered?
  5. (minor) Figure 2 would be clearer without the patterned backgrounds, with larger labels, consistent variable notation, and unified modality representations across Figures 1, 2, and 8.
Reviewer_nYAEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Problem

What is the problem/question?

From the abstract and intro, I can tell that the authors are aiming to produce a more responsive foundation model of multi-modal patient state for pneumonia, but not what that precisely means to them.

Why is the problem/question important/impactful?

Pneumonia is a complex disease and respiratory care is a key area of study in critical care.

Why is the problem/question hard?

Foundation modeling itself is hard, and multi-modal foundation modeling even harder.

What naive solutions have or should have been tried to solve this problem/question?

The naive solution would be to remove the "foundation model" aspect of this, identify the actual key targets of interest, which, in this limited, single-disease context, are probably likely to be not too extensive, and just model them directly with supervised learners.

Beyond naive solutions, there is also an actively developing and developed area of literature on EHR only foundation models that are becoming relatively sophisticated and clearly successful, such as CEHR-GPT, ETHOS/ETHOS-ARES, Curiosity, MOTOR, EveryQuery, and more. I recommend the authors add comments on this literature and why it is or is not relevant to their work.

Why are these naive solutions insufficient?

It is not clear from the text why this is insufficient, nor is any evidence presented (as far as I can tell) that suggests they would not work better.

Solution

How do the authors propose to solve it (in brief)

I can't tell from the paper what their proposed method does. It seems to be a form of a Markov model updating a gaussian representation of belief, combined with diffusion models for estimating out more complex modalities, but I can't tell overall. Significant amounts of notation are introduced without context or sufficient explanation, and the paper is very hard to follow.

Critical strengths

Key Strengths

These are reasons I would advocate the paper be accepted

  1. I think multi-modal health foundation models are an important area of study.

Minor Strengths

These are strengths, but not sufficient alone to warrant advocating acceptance

Critical weaknesses

Key Weaknesses

These are reasons I would advocate the paper be rejected

Your paper has critical clarity and presentation problems that prevent it from being understandable to a reader in this domain without significant intensive effort. You frequently use a lot of "jargon" that makes it hard to parse what claims you're actually making, and fail to provide key details about what you are doing at the appropriate point in the paper. For example, you often use "H1 - H4" in your work, but they are not defined when they are first introduced, instead, you just say "History retention is assessed under a frozen mortality-risk probe using AUROC, AUPRC, and Brier score, with accuracy and balanced accuracy as descriptive metrics. Recursive forecasts advance H1–H4 without factual intermediate observations.". It is not clear what H1-H4 actually are from this, and the only hint of what this means comes 3 pages later, where in a figure caption you say "H1–H4 denote six-hour windows, not acquisition times." Your mathematical notation, starting on page 3, is ambiguously defined and often unclear. E.g., "Let ... denote observed history, including masks, timestamps, and preceding care when available". In what manner does this include "preceding care"? It is not clear what that could mean that would be operationally consistent with your notation. You also need to say what things are -- e.g., when you introduce your belief, what is q theta?

These problems not only make it very challenging to understand what your model is doing, but also how you are judging it against competitors. As I don't know what H1-H4 represent, or how your mortality task is defined, or how you leverage the "three seeds" you reference as you don't report uncertainty, it is very hard to interpret your results tables. You also don't consistently compare to relevant baselines, leaving me unsure what take-aways I should leverage (e.g., your AUROC numbers are without comparators). You need to much more carefully separate out what your various experimental claims / questions are and what evidence you have for those claims.

Minor Weaknesses

These are weaknesses, but not sufficient alone to warrant advocating rejection

  1. I don't like your presentation style for the figures. The hashed backgrounds are distracting and make the figures harder to read.
Questions for authors

Action Items for Rebuttal

Things that, if addressed, may change my score

Unfortunately the clarity issues are too severe for this to be fixed in rebuttal; were the paper presented in clarified form, that would warrant a full new review.

Questions

Things that I want to know, but are unlikely to result in me changing my score

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.