Identification of the Best Anthropometric Predictors of Serum High- and Low-Density Lipoproteins Using Machine Learning

Serum high-density lipoprotein (HDL) and low-density lipoprotein (LDL) cholesterol levels are associated with risk factors for various diseases and are related to anthropometric measures. However, controversy remains regarding the best anthropometric indicators of the HDL and LDL cholesterol levels. The objectives of this study were to identify the best predictors of HDL and LDL cholesterol using statistical analyses and two machine learning algorithms and to compare the predictive power of combined anthropometric measures in Korean adults. A total of 13 014 subjects participated in this study. The anthropometric measures were assessed with binary logistic regression (LR) to evaluate statistically significant differences between the subjects with normal and high LDL cholesterol levels and between the subjects with normal and low HDL cholesterol levels. LR and the naive Bayes algorithm (NB), which provides more reasonable and reliable results, were used in the analyses of the predictive power of individual and combined measures. The best predictor of HDL was the rib to hip ratio (p = <; 0.0001; odds ratio (OR) = 1.895; area under curve (AUC) = 0.681) in women and the waist to hip ratio (WHR) (p = <;0.0001; (OR) = 1.624; AUC = 0.633) in men. In women, the strongest indicator of LDL was age (p = <;0.0001; OR = 1.662; AUC by NB = 0.653); AUC byLR = 0.636. Among the anthropometric measures, the body mass index (BMI), WHR, forehead to waist ratio, forehead to rib ratio, and forehead to chest ratio were the strongest predictors of LDL; these measures had similar predictive powers. The strongest predictor in men was BMI (p = <;0.0001; OR = 1.369; AUC by NB = 0.594; AUC by LR = 0.595). The predictive power of almost all individual anthropometric measures was higher for HDL than for LDL, and the predictive power for both HDL and LDL in women was higher than for men. A combination of anthropometric measures slightly improved the predictive power for both HDL a- d LDL cholesterol. The best indicator for HDL and LDL might differ according to the type of cholesterol and the gender. In women, but not men, age was the variable that strongly predicted HDL and LDL cholesterol levels. Our findings provide new information for the development of better initial screening tools for HDL and LDL cholesterol.