acceptodds
Desk rejected at ICLR 2027

Medical World Regularizer: Learning from Clinical Interventions, Hidden States, and Future Trajectories

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Medical AI systems predict diverse clinical endpoints from patient observations, but clinical care unfolds in a partially observed world shaped by interventions and revealed through subsequent measurements. A native task specifies what should be predicted; it need not specify which aspects of the evolving patient a representation should retain. Our MedWorldReg assigns longitudinal information three roles: past observations describe a state, prior interventions contextualize a decision, and later observations supervise what an earlier state should preserve. Intervention-aware decision making conditions the native readout on observed care. Future-aware state learning predicts representations of subsequent measurements from an earlier state. The two signals meet in a shared temporal encoder, while the task label, loss, and decision time remain unchanged. This task-world coupling uses no clinical simulator or counterfactual rollout, and its future-prediction branch is absent at deployment. The framework applies to longitudinal tasks with the requisite records, without presuming that every endpoint improves.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Biomedical World Models, Joint Embedding Predictive Architecture, Clinical Trajectory Modeling

Related papers

Discussion (0)

Sign in to comment.