Development and Validation of a Prediction Model for Recurrent Atherosclerotic Cardiovascular Disease

BACKGROUND: The definition of very high risk from the American Heart Association/American College of Cardiology 2018 and 2026 cholesterol guidelines may oversimplify heterogeneity in recurrent atherosclerotic cardiovascular disease (ASCVD) risk. OBJECTIVES: We developed a prediction model for recurrent ASCVD and compared its performance with the 2018 cholesterol guideline. METHODS: We used electronic health records from Kaiser Permanente Southern California (KPSC) for model development and the REGARDS (Reasons for Geographic and Racial Differences in Stroke) cohort for external validation. We included adults with a prior ASCVD and followed them for recurrent ASCVD or up to 10 years. We used elastic net regularization for predictor selection and a Fine-Gray subdistribution hazards model for model development. Model performance was evaluated using time-dependent area under the curve, sensitivity, and specificity compared with the 2018 guideline. RESULTS: Among 70,310 and 5,239 adults from KPSC and REGARDS, 17,344 and 1,185 experienced recurrent ASCVD during the median (Q1, Q3) follow-up of 6.1 (2.5, 10) and 8.3 (4.0, 10) years, respectively. Time-dependent area under the curves were 0.68 (95% CI: 0.66-0.69) in KPSC and 0.67 (95% CI: 0.65-0.69) in REGARDS. Compared with the 2018 guideline, using a 5-year 15% risk threshold, the model classified a similar proportion of patients as high risk and improved sensitivity from 64% (95% CI: 62%-66%) to 73% (95% CI: 71%-75%) in KPSC and from 69% (95% CI: 65%-72%) to 73% (95% CI: 69%-76%) in REGARDS; specificity remained similar in KPSC (51% and 52%) and improved in REGARDS (from 45% to 50%). CONCLUSIONS: Compared with the 2018 cholesterol guideline, the new model improves identification of individuals at very high risk of recurrent ASCVD, supporting more personalized treatment decisions.