Testing
Machine learning finds 26-253% more elevated-Lp(a) cases per test than universal screening, model of 438,579 (Eur J Prev Cardiol 2026)
Original title: Machine Learning-driven Prioritisation of Lipoprotein(a) Testing: Model Development and Validation
Development and validation of ethnicity-calibrated machine-learning models to prioritise Lp(a) testing, using UK Biobank participants 37 or older (n=438,579, split into feature-selection, derivation and validation sets) with external validation in NHANES III. Screening one million people with a universal approach would find 222,717 cases above 90 nmol/L and 1950 above 430 nmol/L; targeted ML testing using the same number of tests found 280,899 (+26%) and 6881 (+253%) cases at those thresholds. At intermediate thresholds (125, 150, 175, 200 nmol/L), ML-targeted testing increased yield by 38%, 51%, 59% and 66% respectively. Across the 90-430 nmol/L threshold range, ML-targeted testing (number needed to screen 3.6-145, AUC 0.61-0.84) required 21-72% fewer tests to identify one million cases, with NHANES III showing similar performance; age, height (a sex proxy), total cholesterol and statin use were the top predictors. The authors propose ML-guided prioritisation as a scalable interim step between current low testing rates and universal screening.
Original abstract
Aims: Elevated lipoprotein(a) [Lp(a)] is a common risk factor for cardiovascular disease (CVD) affecting ∼1.4 billion people globally, with novel treatments under development. Guidelines recommend one-lifetime measurement, yet <1% are tested. Population-wide screening faces cost and implementation challenges. We developed a machine learning (ML) model to help prioritise patients for Lp(a) testing.
Methods: Ethnicity-calibrated ML models were developed to identify individuals with elevated Lp(a) in UK Biobank. Participants ≥37 years old (N=438,579) were split into feature importance/selection(20%), derivation(60%), and validation(20%) datasets. Performances across risk-enhancing Lp(a) thresholds recommended by clinical guidelines (90, 125, 430 nmol/L) or entry criteria for ongoing Lp(a)-lowering trials (150, 175, 200 nmol/L) were evaluated. External validation was conducted in NHANES III.
Results: Screening one million people using a universal approach would identify 222,717 cases above 90 nmol/L and 1950 above 430 nmol/L. In contrast, applying ML-targeted testing using the same number of tests would identify 280,899 (+26%; 95%CI:20-28%) and 6881 (+253%; 95%CI:192-310%) cases, respectively. At the thresholds of 125, 150, 175, and 200 nmol/L, yield increases were 38% (95%CI:35-40%), 51% (95%CI:47-54%), 59% (95%CI:55-63%), and 66% (95%CI:61-71%). Across thresholds 90-430 nmol/L, ML-targeted testing (Number Needed to Screen [NNS] 3.6-145, AUC 0.61-0.84) required 21%-72% fewer tests to identify one million cases. NHANES III validation demonstrated similar performance. Top 4 predictors included age, height (proxy for sex), total cholesterol and statin use.
Conclusion: A ML-guided approach to prioritise testing for elevated Lp(a) would require fewer tests to identify those above risk-enhancing thresholds or potentially eligible for emerging therapies, offering a scalable interim compromise between the low current testing rates and universal screening aspirations.
risk predictionscreeningtesting
Summary written by lp-a.org from the published abstract; figures as published. Page updated 17 August 2026. Methods.