Abstract 16483: A New Cardiovascular Disease Risk Model using Historical Repeated Predictors in Electronic Health Records
Introduction: GP-embedded pre-screening to identify high-risk patients for formal risk assessment could be improved by utilizing historical repeated risk predictors. The aim of this study was to develop a new proposed CVD risk algorithm that utilizes historical risk predictors and handles missing data in electronic health records.
Methods: Prospective open cohort study with routinely collected data from across 10 GP practices in England and Wales contributing to The Health Improvement Network database. Data from 63,437 patients aged 40-84 with 3,873 cardiovascular events were split into 2/3 derivation and 1/3 validation cohorts. The main outcome was newly recorded diagnoses of CVD. Risk factors included age, sex, diabetes status, antihypertensive medication use, and all repeated measures of systolic blood pressure, total cholesterol, HDL cholesterol and smoking status. A landmark-age Cox model, stratified by sex, using data from patients without previous CVD and/or statin prescriptions was constructed using summary measures of the historical risk predictors estimated from multivariate mixed models that allow for missing data.
Results: Model fit and risk discrimination from the validation sample indicated that use of historical predictors improves 10-year CVD risk prediction. Although the model complexity increased, model fit was improved with the inclusion of summary statistics from the multivariate mixed models. The C-index, a measure of how well the model discriminates individuals with and without CVD, for a null model with age and diabetes status only was 0.7450 [95% CI: 0.7267, 0.7634] and increased to 0.7667 [0.7490, 0.7844] when repeated measurements of systolic blood pressure, total and HDL cholesterol and smoking status were included in the risk prediction algorithm.
Conclusions: This study outlines an approach to utilizing historical repeat risk predictors in electronic health records for CVD risk prediction, with evidence indicating this approach improves model fit and risk discrimination. Future work will repeat these analyses using data from ~600 GP practices, validate the model against existing CVD prediction models, and evaluate the clinical utility of this algorithm as a pre-screening strategy for primary prevention of CVD.
Author Disclosures: E. Paige: None. J. Barrett: None. I. Nazareth: None. I. Petersen: None. A. Wood: None.
- © 2016 by American Heart Association, Inc.