Prospective Validation of a Machine Learning Model for Low-Density Lipoprotein Cholesterol Estimation.

Abstract

We aim to prospectively validate a previously developed machine learning algorithm for low-density lipoprotein cholesterol (LDL-C) estimation. We retrospectively and prospectively evaluated a machine learning algorithm based on k-nearest neighbors (KNN) according to age, sex, healthcare setting, and triglyceridemia against a direct LDL-C assay. The agreement of low-density lipoprotein-k-nearest neighbors (LDL-KNN) with the direct measurement was assessed using intraclass correlation coefficient (ICC). The analysis comprised 31,853 retrospective and 6599 prospective observations, with a mean age of 54.2 ± 17.2 years. LDL-KNN exhibited an ICC greater than 0.9 independently of age, sex, and disease status. LDL-KNN was in satisfactory agreement with direct LDL-C in observations with normal triglyceridemia and mild hypertriglyceridemia but displayed an ICC slightly below 0.9 in severely hypertriglyceridemic patients and lower in very low LDL-C observations. LDL-KNN performs robustly across ages, genders, healthcare settings, and triglyceridemia. Further algorithm development is needed for very low LDL-C observations.