A Bayesian framework reveals life‑course disease signatures from EHRs and genetics
Researchers introduce ALADYNOULLI, a Bayesian generative model that jointly analyzes longitudinal electronic health records and germline genetics to infer time-varying disease signatures and individual risk trajectories. Validated across UK Biobank, Mass General Brigham, and All of Us (n>683,000) with up to 52 years of follow-up, it identifies 21 stable signatures that align with known biology, enables signature-level GWAS and rare-variant analyses, and supports bias-corrected dynamic risk prediction. The model captures multi-disease progression within a unified framework, allows patient stratification within diagnoses, and demonstrates cross‑cohort replication of signatures while outperforming traditional risk scores for short- and long-term predictions. It also supports practical applications like digital-twin matching and time-sensitive clinical decision support, and its code and data pipelines are publicly available for further validation.