ABSTRACTBackgroundHeart failure (HF) is a major cause of death globally. Prediction of HF risk and early initiation of treatment could mitigate disease progression.ObjectivesThe study aimed to improve the prediction accuracy of HF by integrating genome-wide association studies (GWAS)- and electronic health records (EHR)-derived risk scores.MethodsWe previously performed the largest HF GWAS to date within the Global Biobank Meta-analysis Initiative to create a polygenic risk score (PRS). To extract clinical information from high-dimensional EHR data, we treated diagnosis codes as ‘words’ and leveraged natural language processing techniques to create a clinical risk score (ClinRS). Our method first learned code co-occurrence patterns and extracted 350 latent phenotypes (low-dimensional features) representing EHR codes, then used coefficients from regression of HF on the latent phenotypes in a training set as weights to calculate patient ClinRS in a validation set. Model performances were compared between baseline (age and sex) model and models with risk scores added: 1) PRS, 2) ClinRS, and 3) PRS+ClinRS. We further compared the proposed models with Atherosclerosis Risk in Communities (ARIC) heart failure risk score.ResultsResults showed that PRS and ClinRS were each able to predict HF outcomes significantly better than the baseline model, up to eight years prior to HF diagnosis. By including both PRS and ClinRS in the model, we achieved superior performance in predicting HF up to ten years prior to HF diagnosis, two years earlier than using a single risk predictor alone. Additionally, we found that ClinRS performed significantly better than ARIC model at one year prior to disease diagnosis.ConclusionsWe demonstrate the additive power of integrating GWAS- and EHR-derived risk scores to predict HF cases prior to diagnosis.