一种基于人工智能技术推导低密度脂蛋白胆固醇浓度的新方法

A New Equation to Estimate Low-density Lipoprotein Cholesterol Based on Artificial Intelligence

目的:评估一种基于人工智能技术推导低密度脂蛋白胆固醇(LDL-C)浓度的新方法。方法:收集云南省阜外心血管病医院2017年9月至2021年11月血脂测定数据共118449例样本,整理血脂数据获取特征,构建一种基于人工智能技术推导LDL-C浓度的极限树回归(ETR)模型(LDL-ETR模型)。用LDL-ETR模型的预测值和LDL-C实测值同计算LDL-C浓度的常用公式[Martin/Hopkins公式(LDL-M公式)、Sampson公式(LDL-S公式)、Friedewald公式(LDL-F公式)]进行比较分析。结果:LDL-ETR模型预测值与实测值拟合优度为0.9940,不确定度为12.2109,相关系数为0.9970。当甘油三酯(TG)在0.89~885.11 mg/dl的全浓度范围内,LDL-ETR模型预测值与实测值之间差值为(-0.00±3.50)mg/dl,优于LDL-M公式[(-5.41±7.43)mg/dl]、LDL-S公式[(-6.80±10.91)mg/dl]和LDL-F公式[(-10.06±13.90)mg/dl],P均<0.001;TG对LDL-ETR模型基本无干扰。在测试集总体21398例样本中,LDL-ETR模型中有20101例样本(93.94%)与实测值一致,一致性较好。LDL-ETR模型预测值的逻辑错误率较低,为0.04%,仅次于LDL-M公式的0.02%(P=0.17)。通过学习曲线证明,LDL-ETR模型预测结果适用于相同检验系统的其它样本。结论:这种基于人工智能技术以血脂数据集构建的LDL-ETR模型可较准确地预测LDL-C浓度,相比常用公式,该模型在高或低TG下的预测结果均较好。

Objectives:To establish a new method to predict the concentration of low-density lipoprotein cholesterol(LDL-C)based on artificial intelligence technology using Yunnan population lipid data.Methods:A total of 118449 lipid measurement data of Fuwai Yunnan Cardiovascular Hospital from September 2017 to November 2021 were collected.After data processing to obtain relevant characteristics,an Extra-Trees model(LDL-ETR)was constructed to predict the concentration of LDL-C,which was then compared with measured values and common equations(Martin/Hopkins model,Sampson model and Friedewald model).Results:The LDL-ETR model was more accurate than the other tested equations in terms of predicted value and actually measured value(R^(2)=0.9940,MSE=12.2109,r=0.9970).When the full concentration of triglyceride was in the range of 0.89~885.11 mg/dl,the difference between the value predicted by LDL-ETR model and the actual value was(-0.00±3.50)mg/dl(vs.Martin/Hopkins[-5.41±7.43]mg/dl;vs.Sampson[-6.80±10.91]mg/dl;vs.Friedewald[-10.06±13.90]mg/dl,P<0.001),indicating the negligible interference by triglyceride on predicted LDL-C value with the LDL-ETR model.LDL-ETR model’s predictions of 20101(93.94%)out of the 21398 samples were congruous with the actual values.The logistic error rate of the LDLETR model was also low(0.04%),and is comparable to that of Martin/Hopkins model(0.02%,P=0.17).It was verified by the learning curve that the applicability of the LDL-ETR model could be extendable to other samples of the same test system.Conclusions:This LDL-ETR model constructed with a lipid dataset based on artificial intelligence technology can predict LDL-C concentrations more accurately,and can be applied to samples with both high and low triglyceride concentrations.