In medical diagnosis and treatment, patient states evolve over time while clinical interventions alter future outcomes. Although current medical AI can detect diseases, estimate risks, and generate reports, many systems still return static labels or scores, offering limited insights into how illnesses may progress or how alternative interventions might reshape their trajectories. Medical world models adapt the concept of world models from artificial intelligence to healthcare by learning internal simulators of patient-state dynamics, aiming to help clinicians anticipate deterioration, compare treatment-conditioned futures, and tailor care to individual patients.
Currently, relevant work is scattered across foundation models, longitudinal modeling, disease simulation, treatment-effect estimation, reinforcement learning, and digital twins. To bridge this gap, this review outlines a roadmap for advancing medical AI from isolated diagnosis and prediction toward medical world models that simulate disease evolution and support intervention decisions. This roadmap is organized around three coupled capabilities: patient-state construction, clinical dynamics modeling, and intervention decision support. Through comparisons of representative systems, the contributions of each capability are highlighted, along with how partial components can be integrated into more mature perception-dynamics-planning systems. Finally, the challenges involved in turning plausible rollouts into clinically useful simulators are identified.
Related literature can be found at GitHub.
Blogger's Review: The introduction of medical world models signifies a significant advancement in the field of medical AI. By dynamically simulating patient states, these models can greatly enhance the scientific basis and individualization of clinical decisions, promising impactful roles in disease management and treatment optimization. Looking forward to their practical application and development in the future.