Comparison of a Machine Learning Method and Various Equations for Estimating Low-Density Lipoprotein Cholesterol in Korean Populations

Abstract

Background: LDL-C is the primary target of lipid-lowering therapy and used to classify patients by cardiovascular disease risk. We aimed to develop a deep neural network (DNN) model to estimate LDL-C levels and compare its performance with that of previous LDL-C estimation equations using two large independent datasets of Korean populations.

Methods: The final analysis included participants from two independent population-based cohorts: 129,930 from the Gangnam Severance Health Check-up (GSHC) and 46,470 participants from the Korean Initiatives on Coronary Artery Calcification registry (KOICA). The DNN model was derived from the GSHC dataset and validated in the KOICA dataset. We measured our proposed model's performance according to bias, root mean-square error (RMSE), proportion (P)10-P20, and concordance. P was defined as the percentage of patients whose LDL was within ±10-20% of the measured LDL. We further determined the RMSE scores of each LDL equation according to Pooled cohort equation intervals.

Results: Our DNN method has lower bias and root mean-square error than Friedewald's, Martin's, and NIH equations, showing a high agreement with LDL-C measured by homogenous assay. The DNN method offers more precise LDL estimation in all pooled cohort equation strata.

Conclusion: This method may be particularly helpful for managing a patient's cholesterol levels based on their atherosclerotic cardiovascular disease risk.