Comparison of LDL-C Estimation Using Ridge Regression and Four Established Equations Against Direct Determination of LDL-C in a Northeastern Population in Thailand.

Abstract

Background: Equations traditionally used for estimating low-density lipoprotein cholesterol (LDL-C) have limitations in accuracy and reliability. This study aimed to compare the performance of established equations with a machine learning approach to determine the most appropriate method for LDL-C estimation.

Methods: A retrospective cross-sectional study was conducted using 14,109 lipid profile records from inpatients and outpatients at Kosumphisai Hospital, Northeastern Thailand (2017-2021). LDL-C was estimated using the Friedewald, Puavilai, National Institutes of Health (NIH), and Martin equations, as well as a Ridge regression model. Direct LDL-C measurement served as the reference standard. Model performance was evaluated using mean absolute error (MAE), the proportion of estimates within ±12% of the direct measurement, and Bland-Altman analysis.

Results: The calculation of LDL-C using Ridge regression provided the highest proportion of estimates within the ±12% error margin (75.37%), the lowest MAE (10.05 mg/dL), and the narrowest 95% limits of agreement (-31.19 to 31.57 mg/dL) in Bland-Altman analysis.

Conclusions: Ridge regression provided greater accuracy and reliability for LDL-C estimation compared with the four established equations. Future research should consider incorporating additional predictors and alternative penalized regression techniques, such as Lasso or Elastic Net, to enhance model robustness.