Comparison of conventional formulas and machine learning models for estimating serum low-density lipoprotein cholesterol

Accurate estimation of low-density lipoprotein cholesterol (LDL-C) is essential for cardiovascular risk assessment and monitoring. Because β-quantification is labor-intensive and impractical for routine use, several equations have been developed to estimate LDL-C indirectly, including the widely used Friedewald formula. This study evaluated established LDL-C equations and multiple machine learning models against directly measured LDL-C across different triglyceride and LDL cholesterol levels. A total of 303 serum samples, categorized by triglyceride levels, were analyzed for LDL-C using a direct homogeneous assay on the Roche Cobas® e6000 platform and compared with values estimated by twelve established equations. Agreement was assessed using correlation analysis, mean bias, and Bland–Altman plots. In addition, supervised machine learning models were developed using an 60:40 training–testing split to evaluate predictive performance. All equations except Ahmadi showed strong correlation with direct LDL-C (r > 0.90). The Puavilai and DeLong formulas demonstrated the highest correlation (r ≈ 0.97), with Puavilai showing the lowest mean bias (0.005 mM). Among machine learning models, Artificial Neural Network (ANN), linear regression, and ridge regression achieved the highest correlation (R = 0.97). The ANN showed minimal bias (0.04 mM), while most other models slightly underestimated LDL-C. The Friedewald formula demonstrated increasing positive bias at higher triglyceride levels. The Puavilai formula and ANN model showed superior agreement with direct LDL-C measurements compared with the Friedewald equation. These findings suggest that alternative equations, particularly Puavilai, may improve LDL-C estimation in routine clinical practice in Oman. Further multicenter validation is warranted before national implementation. • LDL-C formulas and machine learning models evaluated against direct LDL-C assay. • Most formulas showed strong correlation with direct LDL-C (r > 0.90). • Puavilai and DeLong formulas showed the lowest bias vs direct LDL-C assay. • ANN, linear, and ridge models showed high predictive accuracy (R ≥ 0.96). • Puavilai formula performed best across varying LDL-C and triglyceride levels.