Understanding how human health evolves and why individuals respond differently to medical interventions is a fundamental challenge in medicine. This paper introduces HealthFormer, a generative AI model designed to act as a "world model" for human physiology. By training on extensive longitudinal data, the model can forecast health trajectories and simulate the potential outcomes of clinical interventions, providing a foundation for developing personalized digital twins. A Generative Approach to Physiology HealthFormer is a decoder-only transformer model trained on data from the Human Phenotype Project, which includes over 15,000 deeply phenotyped individuals. The researchers tokenized health trajectories across 667 distinct measurements. These measurements cover seven diverse domains: blood biomarkers, body composition, sleep physiology, continuous glucose monitoring, gut microbiome, wearable-derived physiology, and behavioral or medication exposure. By training the model to predict these trajectories, the researchers created a system where various clinical tasks—such as risk assessment or intervention simulation—can be performed simply by querying the model. Predicting Health Outcomes A key strength of HealthFormer is its ability to generalize to new populations without requiring task-specific training. When tested on four independent cohorts, the model outperformed established clinical risk scores in predicting 27 out of 30 incident-disease and mortality endpoints. This suggests that the model effectively captures the underlying patterns of human health, allowing it to provide accurate risk stratification across different groups of people. Simulating Clinical Interventions Beyond forecasting, the model demonstrates a unique capability to simulate the effects of medical interventions in silico. In a held-out personalized-nutrition trial, the model successfully predicted individual six-month changes in biomarkers, such as diastolic blood pressure, with high accuracy. Furthermore, when evaluated against 41 randomized intervention-outcome comparisons from published clinical trials, the model correctly predicted the direction of the effect in every instance. In 30 of these cases, the model’s predicted mean outcome fell within the 95% confidence interval reported in the original studies. Toward Clinical Digital Twins The researchers position HealthFormer as an initial step toward creating "clinical digital twins." By treating health as a generative sequence, the model allows clinicians to ask "what if" questions about patient care. Because the model can simulate how a specific individual might respond to different interventions, it offers a promising path toward more personalized, data-driven medicine that accounts for the unique physiological trajectory of every patient.