Estimation of LDL-C using machine learning models and its comparison with directly measured and calculated LDL-C in Turkish pediatric population

Objective: The assessment of lipid profiles in children is critical for the early detection of dyslipidemia. Low-density lipoprotein cholesterol (LDL-C) is one of the most often used measures in diagnosing and treating patients with dyslipidemia. Therefore, accurate determination of LDL-C levels is critical for managing lipid abnormalities. In this study, we aimed to compare various LDL-C estimating formulas with powerful machine-learning (ML) algorithms in a Turkish pediatric population. Materials and Methods: This study included 2,563 children under 18 who were treated at Sivas Cumhuriyet University Hospital in Sivas, Turkey. LDL-C was measured directly using Roche direct assay and estimated using Friedewald's, Martin/Hopkins', Chen's, Anandaraja's, and Hattori's formulas, as well as ML predictive models (i.e., Ridge, Lasso, elastic net, support vector regression, random forest, gradient boosting and extreme gradient boosting). The concordances between the estimates and direct measurements were assessed overall and separately for the LDL-C and TG sublevels. Linear regression analyses were also carried out, and residual error plots were created between each LDL-C estimation and direct measurement method. Results: The concordance was approximately 0.92-0.93 percent for ML models, and around 0.85 percent for LDL-C estimating formulas. The SVR formula generated the most concordant results (concordance=0.938), while the Hattori and Martin-Hopkins formulas produced the least concordant results (concordance=0.851). Conclusion: Since ML models produced more concordant LDL-C estimates compared to LDL-C estimating formulas, ML models can be used in place of traditional LDL-C estimating formulas and direct assays.