预测2型糖尿病并发颈动脉斑块的机器学习模型的建立  

Establishment of machine learning model for predicting carotid plaque in type 2 diabetes mellitus

作　　者：李桃[1,2] 蒋伏松 陈霆[1] 郑西川[1] LI Tao;JIANG Fu-song;CHEN Ting;ZHENG Xi-chuan(Computer Center,Shanghai Jiao Tong University Affliated Sixth People's Hospital,Shanghai 200233,China;Computer Center,Shanghai Sixth People's Hospital East Affiliated to Shanghai University of Medicine&Health Sciences,Shanghai 201306,China;Department of Endocrinology and Metabolism,Shanghai Jiao Tong University Affliated Sixth People's Hospital,Shanghai 200233,China)

机构地区：[1]上海交通大学附属第六人民医院计算机中心,上海200233 [2]上海健康医学院附属第六人民医院东院计算机中心,上海201306 [3]上海交通大学附属第六人民医院内分泌代谢科,上海200233

出　　处：《医疗卫生装备》2020年第1期7-11,24,共6页Chinese Medical Equipment Journal

基　　金：上海经济和信息化委员会信息化发展专项基金资助(201701014);上海健康医学院师资人才百人库项目(2017039)

摘　　要：目的:探索机器学习模型在2型糖尿病患者中预测并发颈动脉斑块的应用价值。方法:使用上海交通大学附属第六人民医院内分泌代谢科(2006—2012年)的住院2型糖尿病实际临床数据集,共8499例患者,约60%为训练集,剩余40%为验证集。采用集成学习算法即随机森林法建立模型分析并验证数十个指标的重要性。挖掘多个检验指标(特征)对颈动脉斑块的影响,并根据这些特征预测颈动脉斑块的有无,为进一步的临床干预或不干预提供依据。为提高颈动脉斑块的预测准确性和效率,剔除一些影响轻微的特征,并根据特征是否易于获取以及重要性逐步减少使用的特征数量。使用相同数据,采用该模型与传统Logistics回归模型进行2型糖尿病并发颈动脉斑块预测。结果:低密度脂蛋白、年龄、空腹胰岛素、甘油三酯、总胆固醇、高密度脂蛋白、空腹血糖、糖化血红蛋白、糖尿病病程均有较高权重,选用这9个特征的机器学习模型具有较高的颈动脉斑块预测精度(80.0%),优于传统的二元Logistics回归模型。结论:在2型糖尿病患者中,使用机器学习模型对是否合并颈动脉斑块有较好的预测效果。Objective To explore the application value of machine learning method in the prediction of carotid plaque in patients with type 2 diabetes mellitus.Methods A total of 8499 patients with type 2 diabetes mellitus(T2DM)from Department of Endocrinology and Metabolism(2006—2012)of Shanghai Jiao Tong University Affliated Sixth People's Hospital were enrolled in the clinical data set,in which about 60%were divided into training sets and the remaining 40%in verification sets.A model was constructed with ensemble learning algorithm(random forests algorithm)to analyze and verify the importance of dozens of indexes.The influences of several test indexes(characteristics)were explored on carotid plaque,and the presence or absence of carotid plaque was predicted according to these characteristics,so as to provide basis for further clinical intervention or nonintervention.In addition,in order to improve the accuracy and efficiency of carotid plaque prediction,some features with slight impact were removed,and the number of features used was gradually reduced according to whether the features were easy to obtain and the doctor's advice.The model developed and Logistics regression model were both used for carotid plaque prediction with the same data sets.Results The experimental results showed that this method not only had a high accuracy of carotid plaque prediction(80.0%),but also could screen the contribution of these characteristics.Low density lipoprotein,age,fasting insulin,triglyceride,total cholesterol,high density lipoprotein,fasting blood glucose,glycosylated hemoglobin Alc,and the course of diabetes all had high weights.Conclusion This method behaves well in carotid plaque prediction for patients with type 2 diabetes mellitus.

关 键 词：2型糖尿病 颈动脉斑块 机器学习 预测 随机森林 数据挖掘 

分 类 号：R318[医药卫生—生物医学工程]