Xian, Su; Smoller, Jordan W.; Luo, Yuan; Walunas, Theresa L.; Liu, Cong; Khan, Atlas; Weng, Chunhua; Kullo, Iftikhar J.; Wei, Wei-Qi; Jarvik, Gail P.; Crosslin, David R. (2026).听.听npj Digital Medicine, 9(1), 682.听
Electronic health records (EHRs) contain large amounts of patient information that can reveal important patterns in health and disease. Using data from more than 100,000 patients, we applied an automated clustering method to group patients based on similarities in their medical conditions, identifying 70 distinct groups. Age and sex strongly influenced these patterns, while the timing of disease onset was closely related to age and risk of death. Genetic differences also contributed to patterns of disease, although their influence appeared to decrease with age. Several patient groups had particularly high levels of multiple health conditions, and these findings were confirmed in an independent dataset. Overall, this study shows how large-scale EHR data can be used to identify meaningful patterns of co-occurring diseases, aging, and health outcomes.
Fig. 1: Illustration of the study framework using 100,272 eMERGE participants EHR data.
a聽Illustration of extracting numerical embeddings from a pre-trained patient embedding model as a representation of digital phenotypes.听b聽Illustration of patient clustering using a Gaussian Mixture model yielding 70 clusters of patients and visualized using T-SNE in two dimensional.听c聽Topic analysis to characterize patient clusters with distinct comorbidity patterns.听d聽Analyzing clusters and phenotype associations with age and sex.听e聽Analyzing cross-phenotype association in cluster groups.听f聽Investigating mortality rate differences in cluster groups and identifying phenotypes associated with premature aging.听g聽Investigation of survival selection and characterizing aging using longitudinal data.