ChronoMedicalWorld: AI Model Predicts Patient Trajectories from EHR Data
A new framework called the ChronoMedicalWorld Model (CMWM) has been introduced by researchers to analyze patient trajectories using longitudinal care data. This model integrates a joint-embedding state encoder with a comprehensive action encoder that incorporates both structured intervention indicators and embeddings from free-text communications. It employs a recurrent latent transition module, guided by a six-term objective that includes next-observation supervision, next-latent prediction, SIGReg latent regularization, and three physiology-aware shape priors (slope, continuity, large-jump penalty). Additionally, a closed-loop rollout-prefix protocol aligns trajectories. CMWM aims to enhance long-horizon clinical simulations for chronic disease management, addressing limitations in current EHR models and general-purpose LLMs. The research is published on arXiv with the identifier 2605.21963.
Key facts
- CMWM is an action-conditioned latent world-model framework
- It learns patient trajectories from longitudinal care data
- Uses joint-embedding state encoder and wide action encoder
- Action encoder handles structured intervention indicators and free-text communication embeddings
- Trains a recurrent latent transition module with six-term objective
- Objective includes next-observation supervision, next-latent prediction, SIGReg regularisation, and three physiology-aware shape priors
- Closed-loop rollout-prefix protocol matches trajectories
- Addresses long-horizon clinical simulation for chronic-disease care
Entities
Institutions
- arXiv