Machine learning-based prediction of LDL cholesterol: performance evaluation and validation.

OBJECTIVE

This study aimed to validate and optimize a machine learning algorithm for accurately predicting low-density lipoprotein cholesterol (LDL-C) levels, addressing limitations of traditional formulas, particularly in hypertriglyceridemia.

METHODS

Various machine learning models-linear regression, K-nearest neighbors (KNN), decision tree, random forest, eXtreme Gradient Boosting (XGB), and multilayer perceptron (MLP) regressor-were compared to conventional formulas (Friedewald, Martin, and Sampson) using lipid profiles from 120,174 subjects (2020-2023). Predictive performance was evaluated using R-squared ( ), mean squared error (MSE), and Pearson correlation coefficient (PCC) against measured LDL-C values.

RESULTS

Machine learning models outperformed traditional methods, with Random Forest and XGB achieving the highest accuracy (  = 0.94, MSE = 89.25) on the internal dataset. Among the traditional formulas, the Sampson method performed best but showed reduced accuracy in high triglyceride (TG) groups (TG > 300 mg/dL). Machine learning models maintained high predictive power across all TG levels.

CONCLUSION

Machine learning models offer more accurate LDL-C estimates, especially in high TG contexts where traditional formulas are less reliable. These models could enhance cardiovascular risk assessment by providing more precise LDL-C estimates, potentially leading to more informed treatment decisions and improved patient outcomes.