AI mines EHR to double the yield of Lp(a) screening

Clinical takeaway: AI applied to electronic health records could help flag more patients for Lp(a) screening as health systems work toward the new universal recommendation. A patient flagged by the model has roughly double the usual odds of an elevated level. 

The first major trial of lowering lipoprotein(a) failed to cut cardiovascular events earlier this month, leaving no approved therapy that targets the particle although others remain in development. But the case for measuring it never really rested on securing a drug to lower it. Lp(a) is a common and genetically determined causal risk factor for atherosclerotic cardiovascular disease (ASCVD) that is largely untouched by lifestyle. A confirmed elevation of Lp(a) marks a patient who warrants more aggressive LDL lowering and tighter control of the other risk factors that treatment can reach. 

The case for testing is now official: this year's US dyslipidemia guideline made a one-time Lp(a) measurement universal for adults. Practice sits far behind, with only about 3% of adults with ASCVD ever tested. A retrospective analysis now supplies the first performance figures for a tool to make screening more systematic: a machine learning model built to identify the ASCVD patients likely to have elevated Lp(a) from EHR data. Prospective validation is under way at large US health systems; the new numbers show what the approach delivered on an earlier test. 

Among adults with ASCVD the model flagged as likely to have elevated Lp(a), 55.1% had a confirmed level of 125 nmol/L or higher, versus 24.8% in the full test population, a 2.2-fold enrichment in screening yield. The model buys that precision by casting a narrow net: it flagged 1,553 of 34,499 adults in the test set and, by design, catches only about 10% of all those with elevated levels. A patient the model passes over has not been screened out. 

Submodels tuned to higher thresholds enriched further, 2.3-fold at 150 nmol/L or higher and 2.7-fold at 200 nmol/L or higher, the level where ongoing trials of Lp(a)-lowering agents have set enrollment. Yield held up across sex and race/ethnicity subgroups: 72% of flagged Black women had confirmed elevations, the highest of any group, against a baseline of 43.6% in that subgroup, itself nearly double the population-wide rate. 

The strongest predictors were use of lipid-lowering medications, led by ezetimibe, rosuvastatin, and evolocumab, followed by diagnosis codes marking prior revascularization and by lipid lab results with higher total cholesterol and LDL paired with lower triglycerides, a profile suggesting the model keys in on patients already on intensified lipid-lowering therapy. Niacin, also among the top medication features, has long been prescribed off-label to lower Lp(a) itself, cutting levels by roughly 20% to 30%. 

The findings come from a retrospective analysis of the Family Heart Database, a national repository of deidentified US medical claims with lab values available in about a third of records. The model, a gradient-boosted classifier using 4,420 features, was trained on 90% of records from 344,987 adults with ASCVD and confirmed Lp(a) levels, then tested on the remaining 10%. It was tuned to favor precision over recall. 

Prospective validation is running now in five US health systems partnered with the Family Heart Foundation, the nonprofit that built the FIND Lp(a) model. That testing is designed to answer the question the authors themselves pose: whether performance holds in patients who have never been tested. The model was trained on adults who already had an Lp(a) result on file, and those patients may differ from untested patients. Clinical use will await results based on use in these health systems' records. 

"Although, recently released U.S. dyslipidemia guidelines recommend Lp(a) screening for all adults, integrating this into routine clinical practice will take many years, if not decades," said Diane MacDougall, MS, vice president of research at the Family Heart Foundation and principal author of the study. "The FIND Lp(a) model supports targeted screening by helping identify people most likely to have high Lp(a), accelerating the adoption of universal screening and creating more opportunities for individuals living with high Lp(a) and their healthcare teams to manage cardiovascular risk." 

Source: MacDougall DE, et al. (2026 Sep 9) JACC Adv. FIND Lp(a) Machine Learning Model: Targeted Screening Enrichment of Elevated Lipoprotein(a) in Atherosclerotic Cardiovascular Disease