Prediction performance of a cardiovascular risk assessment tool using Stanford EHR data repository

Background Stratification of individuals for their risk to develop cardiovascular diseases can be used for effective prevention and intervention. A significant amount of information for risk assessment can be obtained through repurposing electronic health records (EHR). The objective of this study is to derive and assess the performance of prediction models for cardiovascular outcomes by using EHR-derived data. Methods We used the Stanford Medicine Research Data Repository (STARR) data from 2000-2017, containing over 2.1 million patients. A subset of 762,372 individuals with complete International Classification of Diseases (ICD) data was used to fit Cox proportional hazard models for prediction of six cardiovascular-related diseases and type 2 diabetes. Results The derived prediction models indicated consistent high discrimination performance (C-index) for all diseases examined: coronary artery disease (0.85), hypertension (0.82), type 2 diabetes (0.77), stroke (0.76), atrial fibrillation (0.82) and abdominal aortic aneurysm (0.77). Lower prediction abilities were observed for deep vein thrombosis (0.67). These results were consistent across age groups and maintained good prediction abilities among individuals with pre-existing diabetes or hypertension. Assessment of model calibration is ongoing. Conclusions We proposed new prediction models for the seven diseases using ICD codes derived from EHR data. EHR data can be used for health risk assessment, but challenges related to data quality and model generalizability and calibration remain to be solved.