Four Heart Risk Models Face Off in 183,000-Person UK Biobank Test

A head-to-head comparison of four widely used cardiovascular risk prediction models, carried out in nearly 200,000 UK Biobank participants, has found that the older Pooled Cohort Equations outperform the newer PREVENT model in this cohort, with all four tools losing accuracy in older adults. The study, published in BMC Cardiovascular Disorders, evaluated the Pooled Cohort Equations (PCE), the Framingham Risk Score (FRS), Globorisk, and PREVENT for predicting 10-year cardiovascular risk in people without prior cardiovascular disease.

The researchers analyzed 183,124 participants free of cardiovascular disease at baseline and followed them for up to 10 years for incident cardiovascular events. During that period, 10,560 cardiovascular events were recorded, corresponding to a 5.8% event rate. Risk scores for each participant were computed from baseline clinical variables as specified by each model: the 2013 ACC/AHA Pooled Cohort Equations, the 2008 general cardiovascular risk version of the Framingham Risk Score, a UK-specific calibration of Globorisk, and the 2023 American Heart Association PREVENT equations.

Performance was assessed with a battery of standard evaluation methods, including receiver operating characteristic analysis, area under the curve (AUC) with 95% confidence intervals, calibration analysis, sensitivity, specificity, positive and negative predictive values, Decision Curve Analysis, and Net Reclassification Improvement. This combination allowed the team to judge not only how well each model separated people who had events from those who did not, but also how closely its predicted risks matched observed event rates and whether the scores offered practical clinical value at realistic decision thresholds.

The discrimination results were clear-cut. PCE achieved the best overall discrimination, with an AUC of 0.717 (95% CI: 0.712–0.723), narrowly ahead of the Framingham Risk Score at 0.715 (95% CI: 0.709–0.721). PREVENT, the most recent of the four models, landed in an intermediate position with an AUC of 0.668 (95% CI: 0.662–0.674), while Globorisk showed the lowest discrimination at 0.605 (95% CI: 0.599–0.611).

Calibration told a similar story. PCE demonstrated the best calibration across risk strata, meaning its predicted probabilities aligned most closely with the observed proportion of cardiovascular events at different levels of estimated risk. PCE and FRS together demonstrated the best pooled discrimination and calibration among the four models.

Beyond raw statistical performance, the study examined clinical utility. Decision Curve Analysis, which quantifies the net benefit of using a model to guide decisions across a range of risk thresholds, indicated that PCE provided the highest net clinical benefit across most decision thresholds. However, the Net Reclassification Improvement analysis confirmed that no single model provided a consistent reclassification benefit over the others, suggesting that swapping one model for another would not uniformly move patients into more appropriate risk categories.

One of the more striking findings emerged from the sex-stratified analysis. Within each sex, PCE and PREVENT discriminated almost identically: AUCs of 0.672 versus 0.673 in men and 0.709 versus 0.701 in women. The authors conclude that the apparent pooled discrimination gap between these two models is driven substantially by pooling across sexes rather than by a consistent superiority of PCE within either men or women. This nuance matters, because pooled comparisons can mask model behavior that differs by sex, and it suggests PREVENT’s weaker headline performance in this cohort should be interpreted with that composition effect in mind.

Age also emerged as a critical modifier. Model performance declined across all four models in older adults over 59 years of age. The authors argue that this consistent deterioration underscores the need for age-specific recalibration and population-tailored risk prediction frameworks, rather than reliance on any single one-size-fits-all equation.

PREVENT was designed with broader applicability in mind, yet in this cohort it showed intermediate pooled performance, while Globorisk, despite UK-specific calibration, demonstrated the lowest discriminatory ability.

For clinicians and researchers, the practical takeaway is that model choice remains consequential: in this UK Biobank validation, PCE offered the strongest combination of discrimination, calibration, and decision-analytic benefit, but no model dominated on every metric, and all lost accuracy with advancing age.

Subject of Research: Comparative validation of four cardiovascular risk prediction models in the UK Biobank

Article Title: Comparative validation of PCE, Framingham Risk Score, Globorisk, and PREVENT for 10-year cardiovascular risk prediction in the UK Biobank

Article References: N, A., Anith Kumar Jeyalakshmi, S., Vijayan, M., & K Prasad, D. (2026). Comparative validation of PCE, Framingham Risk Score, Globorisk, and PREVENT for 10-year cardiovascular risk prediction in the UK Biobank. BMC Cardiovascular Disorders. https://doi.org/10.1186/s12872-026-06773-x

Image Credits: AI Generated

DOI: 10.1186/s12872-026-06773-x

Keywords: cardiovascular risk prediction, UK Biobank, Pooled Cohort Equations, Framingham Risk Score, Globorisk, PREVENT, model calibration, discrimination, Decision Curve Analysis, primary prevention, age-specific recalibration, Comparative