Among 266,410 UK Biobank participants free of prevalent heart disease, the top-risk 20% flagged by current standard equations — PREVENT (for CHD, HF, stroke) and CHARGE-AF (for atrial fibrillation) — captured only 44.8%–55.2% of subsequent cardiovascular events, revealing a substantial gap in today's prevention targeting. Scanning 3,838 baseline phenotypes and compressing the signal down, researchers identified disease-specific sets of just 13–19 clinically accessible measures that improved held-out AUC by 0.013–0.030 and captured an additional 2.7–4.9 percentage points of events at the same 20% prioritization threshold, with the largest gain for heart failure prediction.
This work is a preprint posted on medRxiv and has not yet been peer-reviewed — conclusions should be treated as provisional. That said, it addresses a genuine and underappreciated problem: most cardiovascular events occur outside the highest-risk stratum, meaning standard triage misses a meaningful share of preventable disease. The approach of distilling thousands of phenotypes into a compact, clinic-ready feature set is pragmatically sound and echoes polygenic-score compression strategies in genomics. Key limitations include the UK Biobank's well-documented healthier-than-average recruitment bias, geographic replication confined within the UK, and the absence of true external validation in diverse populations. The AUC gains, while statistically meaningful at scale, are modest in absolute terms. Whether these additional measures change clinical decisions — and improve outcomes — remains unproven. Incremental rather than paradigm-shifting, but potentially actionable if externally validated.
