A New Low-Density Lipoprotein Cholesterol Estimation
Model from a Linear Regression Model and an Artificial
Neural Network
Thaisiam P, MD¹, Sothornwit J, MD², Charoensri S, MD², Pattanapairoj S, BE³, Kotruchin P, MD, PhD⁴,
Pongchaiyakul C, MD²
Affiliation : ¹ Department of Medicine, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand ² Division of Endocrinology and Metabolism, Department of Medicine, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand ³ Department of Industrial Engineering, Faculty of Engineering, Khon Kaen University, Khon Kaen, Thailand ⁴ Department of Emergency Medicine, Faculty of Medicine, Khon Kaen University, Khon Kaen, Thailand
Background: Low-density lipoprotein cholesterol (LDL-C) estimation from Friedewald equation is frequently used in clinical practice.
However, limitations have emerged regarding its use, including patients with triglyceride (TG) levels of more than 400 milligrams per
deciliters (mg/dL), or LDL-C level of less than 70 mg/dL. Despite that many new LDL-C equation models derived from linear regression
analysis have been proposed, the accuracy of these generated formulas is still questionable. The authors developed a new LDL-C
prediction model constructed by an artificial neural network (ANN), an information processing and computational system modeled
after a biological nervous system, with expected better accuracy than both the Friedewald equation and the linear regression models.
Materials and Methods: A cross-sectional study was conducted. Serum lipid profiles (total cholesterol [TC], TG, high-density lipoprotein cholesterol [HDL-C], and LDL-C) were collected from 10,949 participants irrespective of specimen collection time, time since last caloric intake, comorbidities, and current medications. Direct LDL-C measurement determined by homogeneous assay was considered as the gold standard. Data were randomly divided into two cohorts, one for developing an equation from a linear regression model and ANN model, and another for validation and analyzing the predictive accuracy among the Friedewald equation, linear regression, and ANN model.
Results: The new simple equation derived from the linear regression model was 0.9 TC – 0.1 TG – 0.8 HDL-C. The correlation coefficient between direct LDL-C measurement and Friedewald-calculated LDL-C, LDL-C calculated using linear regression, and ANN-calculated LDL-C were 0.966 (p<0.001), 0.977 (p<0.001), and 0.978 (p<0.001), respectively. The ANN model demonstrated less root mean square error (RMSE) than the Friedewald equation or the linear regression model, which implied better accuracy, even when TG levels were more than 400 mg/dL or direct LDL-C levels were less than 70 mg/dL.
Conclusion: The ANN model is a highly accurate and a better LDL-C estimating tool, even in patients with TG level greater than 400 mg/dL and LDL-C level less than 70 mg/dL.
Received 26 Nov 2019 | Revised 4 Feb 2020 | Accepted 11 Feb 2020
Keywords : Artificial neural network, Friedewald, Equation, Lipid, Low-density lipoprotein cholesterol
All Articles
Download