World Models in Healthcare: From Predictive AI to Clinical Decision Support

Healthcare systems worldwide are grappling with unprecedented s, from aging populations and rising chronic diseases to severe shortages of clinical staff, underscoring an urgent need for AI solutions that go beyond pattern recognition to offer reliable, predictive support for high-stakes medical decisions. Traditional generative models, including large language models and diffusion-based systems, have shown promise in tasks like medical image synthesis and clinical text processing, but they often lack the physical grounding and temporal reasoning required for safe intervention planning, leading to risks like hallucinations in critical scenarios. This gap has spurred interest in world models—AI systems that learn predictive dynamics of clinical environments, enabling simulations of patient states under various actions, which could revolutionize areas from diagnostics to robotic surgery by providing a foundation for counterfactual analysis and planning. As the field evolves, these models aim to shift healthcare AI from static data generation to dynamic, actionable insights, potentially improving outcomes while addressing efficiency pressures in global health systems.

World models in healthcare build on machine learning foundations that emphasize learning transition dynamics, often formalized as predicting future states given current observations and actions, using architectures like the Joint Embedding Predictive Architecture (JEPA) to map latent representations rather than raw data. Researchers leverage large generative backbones such as transformers, diffusion models, and variational autoencoders to handle multimodal inputs—including medical images, electronic health records (EHRs), and genomic data—encoding them into unified latent states where a dynamics predictor models how these states evolve over time or in response to interventions. This ology enables tasks like simulating tumor growth under different treatments or forecasting disease trajectories from EHRs, with a focus on predictive accuracy and the ability to support rollouts for what-if scenarios, moving beyond mere classification or sample generation to embodied, action-conditioned reasoning that mirrors real-world clinical processes.

The review's reveal that current world model applications in healthcare are categorized into three domains: medical imaging and diagnostics, disease progression modeling from EHRs, and robotic surgery and planning, with most systems achieving lower capability levels (L1 temporal prediction and L2 action-conditioned prediction) and fewer advancing to higher tiers like L3 counterfactual rollouts or L4 planning and control. In medical imaging, studies such as CheXWorld use JEPA-style objectives to learn predictive representations from radiographs, while TaDiff employs treatment-aware diffusion models to forecast longitudinal MRI changes in glioma patients under alternative therapies, demonstrating action-conditioned simulation potential. For EHRs, models like Foresight and CoMET utilize generative transformers to auto-regressively predict future medical events from patient timelines, excelling in temporal forecasting but often lacking explicit action semantics, whereas in robotics, systems like EchoWorld and Surgical Vision World Model encode motion-aware dynamics for real-time probe guidance or controllable video generation, showing progress toward closed-loop control in surgical settings.

These advancements carry profound for healthcare, as world models could enable clinicians to simulate treatment outcomes, optimize protocols, and reduce errors through data-driven counterfactual reasoning, potentially accelerating personalized medicine and improving resource allocation in overburdened health systems. By integrating predictive dynamics with causal and mechanistic priors, such models may enhance decision support in areas like oncology, where simulating tumor evolution under various drugs could inform personalized therapy plans, or in surgery, where action-conditioned rollouts might guide robotic instruments with greater precision and safety. However, realizing this potential requires addressing gaps in action specification, interventional validity, and multimodal integration to ensure that simulations translate reliably to clinical practice, fostering trust and adoption among healthcare professionals while mitigating risks associated with AI-driven recommendations.

Despite their promise, world models in healthcare face significant limitations, including underspecified action spaces outside robotics—where variables like drug doses or treatment timing lack formal units and safety constraints—and weak interventional validation that questions the realism of simulated futures in critical scenarios. Many models struggle with incomplete multimodal state construction, failing to fully integrate diverse data sources like imaging, EHRs, and genomics under conditions of missing or irregular sampling, which can compromise the accuracy of rollouts and decision support. Additionally, issues like limited trajectory-level uncertainty calibration and insufficient robustness to out-of-distribution shifts, such as variations in medical devices or patient demographics, highlight the need for improved evaluation standards and causal grounding to ensure these systems are safe, fair, and deployable in real-world clinical environments.

Future research should focus on advancing world models from lower to higher capability levels by formalizing clinical actions, establishing rigorous counterfactual validation through multi-site studies and clinician adjudication, and enhancing multimodal integration and uncertainty quantification to support reliable planning and control. Efforts to embed causal mechanisms and adhere to standardized benchmarks will be crucial for bridging the gap between experimental models and clinical deployment, ensuring that these AI systems can responsibly augment healthcare decision-making without exacerbating existing disparities or safety concerns.