JAMA Netw Open. 2026 Apr 1;9(4):e266838. doi: 10.1001/jamanetworkopen.2026.6838.
ABSTRACT
IMPORTANCE: In 2023, the American Heart Association Cardiovascular-Kidney-Metabolic Scientific Advisory Group introduced the Predicting Risk of Cardiovascular Disease Events (PREVENT) equations, a race-free, sex-specific model for cardiovascular disease (CVD) risk prediction in adults aged 30 to 79 years. While initial validations showed strong performance, their reliability under missingness conditions remains unclear.
OBJECTIVE: To evaluate discrimination and calibration of the PREVENT equations in an electronic health record (EHR) cohort and assess robustness to missingness.
DESIGN, SETTING, AND PARTICIPANTS: This retrospective cohort study used Duke University Health System, a health network encompassing tertiary hospitals, regional hospitals, and primary care practices across North Carolina, EHR data from March 2014 to December 2024 with up to 8 years follow-up. Patients without baseline CVD with sufficient data to calculate PREVENT risk were included. Two cohorts were defined: a relaxed cohort, allowing for missing laboratory and vital sign data with race-sex median imputation, and a strict cohort, restricted to those with complete records. Data were analyzed from October 2024 to June 2025.
EXPOSURES: Published PREVENT equations alongside locally fitted Cox proportional hazards, discrete-time neural network, and recalibrated PREVENT models.
MAIN OUTCOMES AND MEASURES: The primary outcomes were estimated 5-year risk of incident CVD and assessed discrimination (C-index) and calibration (expected vs observed event rates) at 5 years by race, sex, and socioeconomic subgroups. The local adaptation via Duke retraining was compared with machine learning-based recalibration of PREVENT scores.
RESULTS: The study included 406 230 patients in the relaxed cohort (239 764 females with a mean [SD] age of 49 [20] years and 166 466 males with a mean [SD] age of 49 [20] years; 16 291 Asian [4.0%], 107 114 Black [26.4%], and 256 403 White [63.1%]) and 127 151 patients in the strict cohort (71 086 females with a mean [SD] age of 54 [13] years and 56 065 males with a mean [SD] age of 53 [12] years; 8210 Asian [6.5%], 29 033 Black [22.8%], and 83 515 White [65.7%]). PREVENT showed strong discrimination in both cohorts (C-index, 0.77 for both males and females in the strict cohort vs 0.75 for males and 0.77 for females in the relaxed cohort), indicating robustness to missing data. Calibration ratios were higher in the strict cohort, indicating more risk underestimation in the relaxed cohort. Local adaptations minimally affected discrimination and modestly improved calibration.
CONCLUSIONS AND RELEVANCE: In this cohort study, the PREVENT equations showed strong discrimination and generalizability, including with missing laboratory and vital sign data when imputation was applied, supporting reliable CVD risk identification and ranking in routine practice.
PMID:41979878 | DOI:10.1001/jamanetworkopen.2026.6838
