CaMeWM: A Causal Medical World Model for Generative Tumor Evolution and Treatment Planning

TL;DR

CaMeWM is a causal medical world model for support-restricted post-TACE tumor evolution and treatment planning. It forms one intervention-to-decision loop: admit a queried regimen only when fold-local empirical propensity support holds, generate a distribution of lesion-local follow-up worlds via response-grounded anchor plus residual 3D diffusion, then map sampled worlds to RMST using a response-Cox utility model and evaluate the frozen policy with cross-fitted censoring-adjusted doubly robust estimation.

Why it matters

The authors argue that treatment planning is counterfactual and that conditional generation alone can reproduce historical allocation rather than simulate interventional disease dynamics needed to rank regimens. CaMeWM couples an intervention query, a support/positivity restriction, and censoring-aware doubly robust evaluation, aiming to provide an inspectable basis for retrospective, support-restricted treatment planning.

Method

• Supported intervention interface: fix do(A=a) only when a complete regimen has fold-local joint-propensity support, using e_a(h)=P(A=a|H=h) and threshold η.
• Action-conditioned world simulation: generate follow-up worlds using an auxiliary response descriptor, a deterministic anchor preserving stable anatomy, and residual 3D diffusion transition (lesion-local residual variation).
• Utility-aware rollout and evaluation: map factual or sampled states to restricted mean survival time (RMST) with a response-Cox model; rank supported actions via generated-world RMST and evaluate with cross-fitted censoring-adjusted doubly robust estimation.

Limitation

The authors state that the learned networks cannot verify causal assumptions (e.g., consistency, conditional exchangeability, positivity); retrospective evaluation does not establish prospective clinical benefit, and they do not claim a unique alternate-world CT for an individual patient.

Related papers

• CoRe-WM: Conditioned Response-State World Modeling for Longitudinal Treatment-Response Prediction in Breast DCE-MRI
• PatientWM: Towards Generalizable Clinical World Models via Unified Patient States
• Same Predictions, Different Harms: Causal Auditing of Patient World Models
• Patient, Place, Prior (P$^3$): What Counts as Personalization in Medical World Models?
• Repeated-Measure Leakage, Distribution Shift, and Reliability under Partial Observation in Patient World Models