A world model simulates the latent dynamics of human health

HealthFlux is a new world model that learns to predict human health status based on 5,647 features from eleven data domains including clinical records, blood tests, genetics, proteomics, metabolomics, and MRI. The model was developed on a sample of 502,166 UK Biobank participants. HealthFlux can predict 195 diseases and death within five years with an average accuracy of 0.816, significantly better than the previous best model with accuracy of 0.715. The model was validated on three independent cohorts and outperforms specialized clinical risk scores. Notably, the model can predict diseases it was not trained on, suggesting it learned health itself rather than specific diseases. Each data modality contributes information that others lack, and their integration identifies at-risk patients that single-modality models would miss.