Comparison of a Machine Learning Method and Various Equations for Estimating Low-Density Lipoprotein Cholesterol in Korean Populations

Abstract

Background: LDL-C is the primary target of lipid-lowering therapy and used to classify patients by cardiovascular disease risk. We aimed to develop a deep neural network (DNN) model to estimate LDL-C levels and compare its performance with that of previous LDL-C estimation equations using two large independent datasets of Korean populations.

Methods: The final analysis included participants from two independent population-based cohorts: 129,930 from the Gangnam Severance Health Check-up (GSHC) and 46,470 participants from the Korean Initiatives on Coronary Artery Calcification registry (KOICA). The DNN model was derived from the GSHC dataset and validated in the KOICA dataset. We measured our proposed model's performance according to bias, root mean-square error (RMSE), proportion (P)10-P20, and concordance. P was defined as the percentage of patients whose LDL was within ±10-20% of the measured LDL. We further determined the RMSE scores of each LDL equation according to Pooled cohort equation intervals.

Results: Our DNN method has lower bias and root mean-square error than Friedewald's, Martin's, and NIH equations, showing a high agreement with LDL-C measured by homogenous assay. The DNN method offers more precise LDL estimation in all pooled cohort equation strata.

Conclusion: This method may be particularly helpful for managing a patient's cholesterol levels based on their atherosclerotic cardiovascular disease risk.

Keywords: Korean; cardiovascular disease; deep neural network; low-density lipoprotein; pooled cohort equation.

Figure 1

Conceptual schematic for internal and external validation of machine learning models. For internal validation, the Gangnam Severance Health Check-up (GSHC) was reserved for testing the model performance. For external validation, Korean Initiatives on Coronary Artery Calcification (KOICA) registry was used to test the model performance. The DNN consists of six hidden layers, four hidden layers, and two hidden layers, with 30 nodes in each layer. Ten cross-validation was performed to determine the structure of DNN in the derivation set. We selected the best DNN model with lowest mean standard error (MSE) among the three layers. A total of three DNN models were competed. The final model was validated using the external validation set.

Figure 2

(A,B) DNN model selection. The DNN consists of six hidden layers, four hidden layers, and two hidden layers, with 30 nodes in each layer. Black bars, upper margins, and maximum lines in each boxplot indicate the means, one standard deviation (SD), and 1.96 SD deviation values, respectively. The best DNN model among each layer was selected, and the tournament method was used to identify the final model. The 30, 30 in the two-layer model had the lowest MSE and was selected as the final model.

Figure 3

Clinical characteristics of the study population among the three datasets. P values were calculated using the t-test and Mann-Whitney U test for continuous variables or the chi-square test for categorical variables. ***p < 0.001. PCE, pooled cohort equations.

Figure 4

Lipids profiles and LDL values from the various LDL equations. The distribution of whole dataset was described. P values were calculated using the t-test and Mann-Whitney U test for continuous variables.

Figure 5

Performances of four LDL estimation methods. (A) Bias of four LDL estimation models. (B) RMSE of four LDL estimation model. (C) P10, 15, and 20 of four LDL estimation models. (D) Concordances of four LDL estimation models. DNN, deep neural network; FW, Friedewald LDL-C; LDL, Low-density lipoprotein cholesterol; NIH, NIH equation for LDL-C; Novel, novel equation for LDL-C.

Figure 6

Comparison of root mean-square errors (RMSEs) for four LDL estimation models. (A) RMSEs for four LDL estimation models according to triglyceride classification. (B) RMSE for the four LDL estimation models according to non-HDL cholesterol classification.

Figure 7

Comparison of RMSE in the four LDL estimation models across the PCE categories. PCE, pooled cohort equations. (A) PCE scores were classified based on the 20th decile. (B) The RMSEs of each LDL-C estimation methods were separately calculated according to each of the 20 PCE categories.