medDreamer
Summary
medDreamer is an action-conditioned latent world model and offline reinforcement-learning framework for irregular clinical time series. It combines missingness-aware Adaptive Feature Integration, recurrent prior/posterior patient-state dynamics, and a two-phase real-grounded-to-imagined policy curriculum.
Role In The Wiki
medDreamer is the wiki’s first direct numeric clinical example of a learned world model trained from irregular multivariate observations and logged interventions. It complements MIRA: MIRA is a passive irregular-time forecasting foundation model, while medDreamer explicitly conditions latent dynamics on treatment actions and optimizes a policy through imagined trajectories.
The distinction is also its main caveat. Historical treatments are confounded logged decisions, so the system demonstrates an action-conditioned modeling interface without establishing causal counterfactual validity or real patient benefit.
Evidence
Official Artifacts
Artifact caveat: the code release contains the core implementation but no EHR preprocessing pipeline, checkpoints, tests, or complete locked environment. Its README command and parser disagree, and GitHub has no detected repository license even though Zenodo labels the archived release CC BY 4.0.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in meddreamer-2026 rather than duplicating verdict rows here. At the entity level, medDreamer contributes a concrete interface:
irregular multivariate observations + masks + elapsed time + logged interventions
-> latent patient state
-> action-conditioned imagined trajectories
-> offline policy optimizationThis is direct time-series/world-model evidence, but the retrospective clinical setting leaves causal identification, calibrated uncertainty, policy overlap, prospective safety, and cross-institution transfer unresolved.