Extended Data Fig. 3: Comparative performance of CAD risk prediction models in the UK Biobank.

This figure compares the predictive performance of our final perspective 10-year CAD meta-prediction model to established clinical risk scores, including PCE, QRISK3, PREVENT, and previous polygenic score benchmarks including GPS_CAD, metaGRS_CAD, Aragam₂₀₂₂, Tcheandjieu₂₀₂₂, and GPS_Mult, as well as ML models including ML4H_EN-COX and UKBCRP using the incident CAD cohort (n = 33,419). For each model, the left panel depicts a scatter plot illustrating the incidence of CAD events across percentile bins of predicted risk, showcasing the predictive density of each model at 10-years of follow-up, while the right panel displays the 10-year cumulative risk trajectories for each risk prediction model, highlighting the ability of each model to stratify risk across time. Shaded regions in the right panel represent 95% confidence intervals (CI). In all cases, meta-prediction significantly outperforms prior approaches.