Hongtu Zhu, PhD

Kenan Distinguished Professor of Biostatistics, Gillings School of Global Public Health

Multimodal disease routes link risk to biology across the human phenome

Electronic health records enable prediction across the disease spectrum, while biobank studies connect molecular and organ phenotypes to disease. Connecting these advances requires representations that preserve how different evidence sources contribute to an individual’s predicted risk. Here we introduce RouteDx, which represents disease prediction scores as additive, patient-specific routes organized along the diagnosis hierarchy. Two variants share a clinical backbone and combine genetic information with proteomic or organ-imaging measurements. In 502,167 UK Biobank participants, RouteDx predicted 1,335 incident diagnoses and outperformed existing longitudinal predictors and cardiovascular risk equations in the evaluated benchmarks. Route contributions form part of the predictive computation, allowing each score to be traced to its constituent evidence sources. Within diagnoses, route profiles identified patient strata whose phenotypic contrasts were preserved in held-out participants, exceeding stratified permutation expectations for 46 of 63 diseases. A separate analysis found concordant disease–measurement associations in UK Biobank and MIMIC-IV, providing external evidence for shared phenotypic relationships. These findings establish multimodal disease routes as a representation connecting phenome-wide prediction with patient-level variation in clinical and biological evidence. This representation enables investigation of heterogeneity within diagnoses and generates hypotheses about the biological differences accompanying distinct patterns of predicted risk.