A generative digital twin forecasts and simulates intensive care patients without task labels

Abstract Generative models that simulate a patient's futures could serve as digital twins for prognosis, trial design and teaching, but in intensive care they have been trained within single databases and rarely tested against observed treatment responses. We trained Didymos, a 123-million-parameter transformer, without task labels on 409,297 stays from seven databases in five countries. On held-out stays, Didymos forecast 13 primary endpoints with AUROCs of 0.83–0.98 from its simulated futures and ranked first among 20 predictors, including supervised LightGBM and TabPFN-3.5 (post hoc). Its step hazards of death were calibrated, whereas free generation postponed deaths. Simulated treatment responses agreed with eight of nine observational contrasts, membership inference stayed near chance (AUC ≤ 0.57) and raters did not tell synthetic from real patients. Didymos risks cut simulated trial sizes by 14–37%. At two unseen hospitals, minutes of training on unlabelled local stays brought Didymos level with locally supervised models.