PREVENT Equations Show Strong CVD Risk Prediction in EHR Data Despite Missing Information

The Predicting Risk of CVD Events (PREVENT) equations, which have been introduced to modernize cardiovascular risk prediction, demonstrated strong performance in real-world electronic health record (EHR) settings—even when key clinical data were missing—supporting their broader use for identifying patients at elevated cardiovascular disease (CVD) risk in routine clinical practice.¹

This retrospective cohort study is published in JAMA Network Open.

“These findings also underscore the importance of evaluating models in the settings and populations where they will be used,” wrote the researchers of the study. “In practice, PREVENT is most useful for population-level risk stratification to guide targeted outreach, follow-up, or referral and may be especially valuable in primary care and large systems with missing or inconsistently collected measurements.”

A prior analysis of the PREVENT equations demonstrated strong performance in predicting 10-year CVD risk across diverse racial and ethnic groups, reinforcing its role as a modern, race-free risk prediction tool.² In that study, PREVENT maintained good discrimination and calibration overall, although performance varied modestly across disaggregated subgroups.

The current study used EHR data from the Duke University Health System spanning March 2014 to December 2024, with up to 8 years of follow-up.¹ Adults aged 30 to 79 years without baseline CVD were included and stratified into 2 cohorts: a “strict” cohort with complete clinical data and a larger “relaxed” cohort that allowed for missing laboratory and vital sign data, which were addressed using race- and sex-specific median imputation.

Investigators evaluated the performance of the published PREVENT equations alongside locally adapted models. Model performance was assessed based on discrimination using the C-index and calibration by comparing expected vs observed 5-year cardiovascular event rates across demographic and socioeconomic subgroups.

The study included 406,230 patients in the relaxed cohort and 127,151 patients in the strict cohort. In the relaxed cohort, 239,764 were female with a mean (SD) age of 49 (20) years and 166,466 were male with a mean age of 49 (20) years; the cohort was racially and ethnically diverse, including 16,291 Asian patients (4.0%), 107,114 Black patients (26.4%), and 256,403 White patients (63.1%).

The strict cohort included 71,086 females with a mean age of 54 (13) years and 56,065 males with a mean age of 53 (12) years, with 8210 Asian (6.5%), 29,033 Black (22.8%), and 83,515 White (65.7%) patients.

The PREVENT equations demonstrated strong and consistent discrimination across both cohorts, with C-index values of 0.77 for males and females in the strict cohort and 0.75 for males and 0.77 for females in the relaxed cohort, indicating robust performance despite missing data. Calibration analyses showed higher calibration ratios in the strict cohort, suggesting greater risk underestimation in the relaxed cohort with imputed data. Local model adaptations produced minimal changes in discrimination and only modest improvements in calibration.

However, the researchers noted some limitations. Follow-up was limited to a maximum of 8 years, restricting long-term risk assessment and preventing full 10-year evaluation, although results were consistent at 5 and 8 years. Additionally, the analysis was conducted within a single North Carolina health system, which may limit generalizability and could miss events occurring outside the network. Furthermore, the researchers did not evaluate time-updated models. Lastly, use of race- and sex-specific median imputation may not fully capture complex data relationships, and more advanced imputation methods may further improve calibration in future work.

Despite these limitations, the researchers believe these findings support the utilization of the PREVENT equations for identifying and stratifying CVD risk in real-world EHR settings, including in the presence of missing data.

“Overall, the PREVENT equations appear well suited for routine care, even with incomplete data, given consistent discrimination across diverse subgroups,” wrote the researchers. “However, calibration disparities in disadvantaged populations suggest health systems should perform subgroup-specific calibration checks and consider recalibration before deployment, particularly when absolute risk thresholds guide treatment.”

References

1. Hong C, Niu M, Wang H, et al. Performance of PREVENT cardiovascular risk in electronic health record–based clinical practice. JAMA Netw Open. 2026;9(4):e266838. doi:10.1001/jamanetworkopen.2026.6838

2. Steinzor P. PREVENT equations accurately predict 10-year CVD risk across diverse groups. AJMC^®. February 12, 2026. Accessed April 15, 2026. https://www.ajmc.com/view/prevent-equations-accurately-predict-10-year-cvd-risk-across-diverse-groups