Among 438,640 UK Biobank participants free of cardiovascular disease at baseline, 45,003 incident CVD events occurred over 13.5 years of follow-up. Head-to-head validation of 16 risk-prediction models across 19 configurations found dramatic performance variation: 10-year AUC ranged from 0.668 (QRISK3) to 0.734 (PREDICT), while calibration failures were widespread — Framingham, PROCAM, ASSIGN, and QRISK1 systematically overestimated risk, whereas PREVENT, SCORE2, and SCORE2-OP underestimated it. Composite assessment favored PREVENT, QRISK2, PREDICT, PCE, and northern China-PAR variants.
The implications here are substantial. Cardiovascular risk scores underpin statin prescribing, blood pressure treatment thresholds, and preventive counseling for hundreds of millions of adults worldwide. Framingham, one of the most globally embedded models, emerged among the poorer performers — a finding consistent with prior single-model external validations but now contextualized against 15 competitors simultaneously. The divergence in net benefit at higher treatment thresholds is particularly clinically meaningful: choosing the wrong model could translate to either over-medicating low-risk individuals or under-treating those who would genuinely benefit.
Limitations are notable: the UK Biobank skews toward healthier, predominantly white British participants, limiting generalizability across ethnicities and lower socioeconomic groups. The study is also observational and cannot adjust for post-baseline treatment changes. As a preprint not yet peer-reviewed, these rankings may shift following expert scrutiny. Nevertheless, this represents a rare, methodologically rigorous simultaneous benchmark — an incremental but practically important advance for clinicians selecting which score to deploy locally.