Transferability and accuracy of electronic health record-based predictors compared to polygenic scores

Electronic health record (EHR)-based phenotype risk scores (PheRS) leverage individuals’ health trajectories to infer disease risk. Similarly, polygenic scores (PGS) use genetic information to estimate disease risk. While PGS generalizability has been previously studied, less is known about PheRS transferability across healthcare systems and whether PheRS provide complementary risk information to PGS.

We trained PheRS to predict the onset of 13 common diseases with high health burden in a total of 845,929 individuals (age 32-70) from 3 biobank-based studies from Finland (FinnGen), the UK (UKB) and Estonia (EstB). The PheRS were based on elastic-net models, incorporating up to 242 diagnoses captured in the EHR up to 10 years before baseline. Individuals were followed up for a maximum of 8 years, during which disease incidence was observed. PGS were calculated for each disease using recent publicly available results from genome-wide association studies.

All 13 PheRS were significantly associated with the diseases of interest. The PheRS trained in different biobanks utilized partially distinct diagnoses, reflecting differences in medical code usage across the countries. Even with the large variability in the prevalence of various diagnoses, most PheRS trained in the UKB or EstB transferred well to FinnGen without re-training. PheRS and PGS were only moderately correlated (Pearson’s r ranging from 0.00 to 0.08), and models including both PheRS and PGS improved onset prediction compared to PGS alone for 8/13 diseases. PheRS was able to identify a subset of individuals at high-risk better than PGS for 8/13 disease.

Our results indicate that EHR-based risk scores and PGS capture largely independent information and provide additive benefits for disease risk prediction. Furthermore, for many diseases the PheRS models transfer well between different EHRs. Given the large availability of EHR, PheRS can provide a complementary tool to PGS for risk stratification.