medDreamer

Source

Raw Markdown: paper_meddreamer-2026.md
PDF: paper_meddreamer-2026.pdf
Latest preprint: arXiv 2505.19785v3
Publisher record: ACM DOI 10.1145/3770854.3780292
Official code: xuqianyi/medDreamer
Archived code release: Zenodo 10.5281/zenodo.18041379
Local code-artifact audit: papers/meddreamer-2026/code_artifact_audit.json

Status And Credibility

The preprint was first submitted on 2025-05-26 and revised to v3 on 2025-12-02. ACM records the paper as a peer-reviewed proceedings article in KDD 2026, published online on 2026-04-20 and in the conference proceedings on 2026-08-09, pages 1693–1704. The authors are Qianyi Xu, Gousia Habib, Feng Wu, Dilruk Perera, and Mengling Feng at the National University of Singapore.

This is current, credible evidence rather than an old architecture-only source: KDD is a tier-1 data-mining venue, the 2026 proceedings record is authoritative, and an official implementation plus a Zenodo-pinned v1.0.0 archive exist. No verified official project page, lab/author blog, Hugging Face artifact, or official X announcement was found during ingest; an authenticated exact-title X search returned no results.

Core Claim

medDreamer is a non-vision, action-conditioned latent world model for irregular clinical time series. It encodes observed values together with per-feature elapsed times, rolls a latent patient state forward under logged treatment actions, and trains an actor–critic first on real-grounded hybrid trajectories and then on fully imagined trajectories.

The paper shows a concrete route from irregular multivariate EHR observations to latent imagined rollouts. It does not establish that the imagined trajectories are causal counterfactuals or that the recommended treatments improve real patient outcomes: all policy evidence comes from retrospective observational data and off-policy proxy evaluation.

Model Interface

flowchart LR
  O["irregular vitals/labs + mask"]
  DT["per-feature elapsed time"]
  A["logged treatment action"]
  AFI["Adaptive Feature Integration"]
  Post["posterior latent state"]
  Prior["action-conditioned prior dynamics"]
  Heads["reconstruction + reward + continuation"]
  P1["phase 1: 50 real-grounded + 10 imagined steps"]
  P2["phase 2: 15-step imagined rollouts"]
  AC["actor–critic policy"]

  O --> AFI
  DT --> AFI
  AFI --> Post
  A --> Post
  Post --> Prior
  A --> Prior
  Post --> Heads
  Prior --> Heads
  Heads --> P1 --> AC
  P1 --> P2 --> AC

Irregular-observation encoder

At time $t$ , the model receives numeric observations $o_{t}$ , a binary observation mask $m_{t}$ , and a per-feature elapsed-time vector $Δ_{t}$ . The Adaptive Feature Integration (AFI) module linearly embeds observed values and elapsed times, then uses a factorization-machine interaction term to model pairwise cross-feature structure. Missingness timing is treated as input signal rather than filled with a fixed mean or median.

Latent dynamics

A posterior GRU infers latent state from the prior state, previous action, and AFI observation representation. A prior GRU predicts the next latent state from state and action alone so trajectories can continue without future observations. Reward, continuation, and masked reconstruction heads train the latent state to retain decision- and observation-relevant information.

Two policy phases

Clinical grounding: each training trajectory uses $T = 50$ posterior states from real observations and logged clinician decisions, followed by $τ = 10$ imagined states. The actor and critic train on the concatenated latent trajectory.
Longer imagination: the policy is refined on fully imagined rollouts of horizon $H = 15$ .

This is a curriculum over the source of latent states, not a formal safety guarantee. The paper’s phrase “safe exploration” means offline rollout constrained by a learned model and historical grounding; no certified constraint, uncertainty shield, prospective trial, or live safety intervention is evaluated.

Data And Action Semantics

Task	Dataset and cohort	Observations	Action / intervention interface	Outcome used by the paper
Sepsis treatment	MIMIC-IV, 21,233 adult septic patients	40 demographics, vitals, scores, labs, and treatment-status features	25 discrete combinations from five IV-fluid bins × five vasopressor bins	90-day mortality plus shaped SOFA/lactate rewards
Mechanical ventilation	eICU-CRD, 21,595 ventilated patients	41 demographics, vitals, labs, medication, and ventilation-related features	18 discrete combinations from PEEP, FiO₂, and tidal-volume bins	Hospital mortality plus shaped SpO₂/mean-blood-pressure rewards
Renal replacement therapy appendix	AmsterdamUMCdb	Appendix-only evaluation details	Dialysis necessity/mode decisions	OPE metrics only in the reported appendix table

These are logged clinical decisions in observational EHRs. They are actions or interventions in the model interface, but their historical assignment is state-dependent and confounded. A learned transition model can simulate conditional trajectories without identifying the causal treatment response.

Evidence And Results

The main table reports the following point estimates:

Task / policy	WIS ↑	WPDIS ↑	CWPDIS ↑	Estimated mortality ↓	Reported 95% CI
Sepsis / clinician	4.87	0.14	0.09	23.44%	[22.12, 24.77]
Sepsis / medDreamer	9.25	0.27	0.19	14.25%	[14.6, 17.5]
MV / clinician	13.31	0.24	0.18	18.27%	[15.48, 21.06]
MV / medDreamer	24.06	0.32	0.25	12.44%	[11.42, 13.45]

Other positive evidence includes:

AFI, the two policy phases, imagination lengths, recurrent backbone choice, and added random missingness are ablated.
The paper reports stable sepsis OPE metrics through 20% additional random masking, with larger degradation at 50%.
Reconstructed mean-blood-pressure examples track observed trajectories more closely than the Dreamer baseline, especially under high missingness.
The appendix reports better OPE values than clinician and Dreamer baselines on an AmsterdamUMCdb renal-replacement-therapy task.
The action-distribution plots show severity-aware policy changes, but these are descriptive policy comparisons rather than evidence of superior clinical care.

Numerical And Evaluation Audit

Two internal numerical inconsistencies materially limit the mortality headline:

The sepsis medDreamer point estimate is 14.25%, but Table 1 reports the 95% CI as [14.6, 17.5]; the point estimate lies outside the interval. The following prose instead gives [14.6, 15.7], which still excludes 14.25% and disagrees with the table.
The contribution list claims mortality decreases of 11.19% and 5.83% versus clinicians. The table values imply 9.19 percentage points for sepsis ( $23.44 - 14.25$ ) and 5.83 percentage points for MV ( $18.27 - 12.44$ ). The sepsis claim is two points larger than the table arithmetic; the relative reductions would instead be 39.21% and 31.91%.

The mortality values are not observed outcomes under deployed policies. The paper estimates them by relating expected returns to observed mortality under clinician actions. The OPE stack also depends on an LSTM approximation of clinician behavior propensities. The methods name effective sample size (ESS), but the main results table does not report ESS or a policy-overlap diagnostic. These limitations make WIS/WPDIS/CWPDIS and estimated mortality useful retrospective signals, not prospective evidence of treatment benefit.

Official Code Audit

The official repository contains the core Python model and configuration files and is archived as Zenodo v1.0.0. The public artifact is not an end-to-end reproduction package:

no MIMIC-IV/eICU extraction or preprocessing pipeline, trained checkpoints, evaluation outputs, tests, or environment lockfile are present;
the README documents --config, while main.py defines --configs;
main.py imports scikit-learn, but requirements.txt omits it;
GitHub exposes no repository license, while Zenodo labels the archived release CC BY 4.0. Reuse terms therefore need clarification rather than inference from the Zenodo metadata alone.

Limitations And Gotchas

All experiments are retrospective and observational; confounding, behavior-policy estimation, overlap, and reward-model bias remain.
No prospective trial, external clinical deployment, or expert blinded comparison tests whether recommendations improve outcomes or are acceptable to clinicians.
The learned world model and reward head can be exploited by the policy. Reconstruction quality and OPE do not establish accurate long-horizon treatment-response counterfactuals.
The paper reports task-specific seeds and one 80/20 patient split, but no multi-seed policy uncertainty or cross-hospital validation for the headline results.
The two-phase curriculum is empirically motivated, not a certified safety mechanism.
The main result’s sepsis confidence interval and mortality-reduction arithmetic are internally inconsistent.
Generalization beyond the selected cohorts, action discretizations, reward definitions, and institutions is unproven.

Foundation TSFM Relevance

Agenda slot	Verdict	Evidence	Missing pieces
Time representation and irregular event streams	partially closes	Uses per-feature elapsed times and masks instead of coarse fixed-window imputation for irregular EHR observations.	Still uses task-specific curated feature tables; no general event schema, continuous-time benchmark, or always-on stream.
Latent-state prediction	partially closes	Posterior/prior recurrent dynamics maintain and roll forward a latent patient state with reconstruction, reward, and continuation heads.	No latent-state ground truth, calibrated predictive belief, rare-state probes, or multi-institution transfer.
Native multivariate encoding	partially closes narrowly	AFI models interactions among 40–41 clinical features while retaining missingness timing.	Low-channel healthcare tasks only; no arbitrary channel schema or high-dimensional scaling.
Control and counterfactuals	partially closes for interface, warning for validity	Logged treatments are explicit action channels and the model performs action-conditioned imagined rollouts for policy learning.	Observational confounding, estimated behavior propensities, learned reward bias, no counterfactual ground truth, and no prospective policy transfer.
Benchmark and evaluation hygiene	warning	Reports several OPE estimators, mortality–return plots, ablations, missingness stress tests, and an extra dataset/task.	Missing ESS/overlap reporting, no multi-seed uncertainty, no live outcomes, incomplete public reproduction stack, and internal numerical inconsistencies.

Links Into The Wiki

Open Questions

Can a confounding-aware, uncertainty-calibrated patient world model preserve medDreamer’s irregular-time advantages without turning imagined trajectories into unsupported causal claims?
Do AFI gains persist against modern irregular-time baselines under identical preprocessing, action definitions, and compute?
How do ESS, behavior-policy overlap, and alternative OPE estimators change the headline ranking?
Which errors come from latent dynamics, reward modeling, behavior-policy estimation, and mortality calibration respectively?
Can the two-phase real/imagined curriculum improve a simulator-backed or randomized treatment benchmark where action consequences are identifiable?
Will expert review and prospective silent deployment agree with the retrospective action-distribution and estimated-mortality signals?

Alex Open Research Wiki

Explorer

medDreamer: Model-Based Reinforcement Learning with Latent Imagination on Complex EHRs for Clinical Decision Support