FIND Lp(a) Machine Learning Model: Targeted Screening Enrichment of Elevated Lipoprotein(a) in Atherosclerotic Cardiovascular Disease.

BACKGROUND: The Flag, Identify, Network, Deliver (FIND) lipoprotein(a) (Lp[a]) machine learning model (MLM) is part of a Family Heart Foundation collaborative quality improvement initiative intended to accelerate Lp(a) screening adoption in U.S. health care systems. OBJECTIVES: To train, test, and characterize important features of an MLM that uses deidentified medical records to flag adults with atherosclerotic cardiovascular disease (ASCVD) with likely elevated Lp(a). METHODS: A Light Gradient-Boosting Machine MLM was trained (90% of data) and tested (10% of data) in 344,987 adults with ASCVD in the Family Heart Database. Percentage of adults with confirmed Lp(a) ≥125 nmol/L were compared (test data set vs cohort flagged by model), including submodels (Lp[a] ≥150 nmol/L, ≥200 nmol/L). The importance of 8 feature groups to MLM performance was determined using Tree Shapley Additive exPlanations. RESULTS: The test data set (n = 34,499) included 8,556 (24.8%) adults with confirmed Lp(a) ≥125 nmol/L. FIND Lp(a) MLM flagged 1,553 adults, 856 (55.1%) had confirmed Lp(a) ≥125 nmol/L, representing a 2.2-fold screening enrichment due to the model (55.1% vs 24.8%). Supportive ≥150 nmol/L and ≥200 nmol/L submodels had 2.3 and 2.7 fold screening enrichment, respectively. Medication use was the most important feature group followed by diagnoses and lipid labs. CONCLUSIONS: One-half of adults with ASCVD flagged by FIND Lp(a) MLM had confirmed elevated Lp(a) ≥125 nmol/L. A 2.2-fold screening enrichment was observed (≥125 nmol/L model); enrichment was greater with ≥150 and ≥200 nmol/L submodels. Early performance indicators support the central role of FIND Lp(a) MLM within the ongoing FIND Lp(a) program.