基于人工智能的糖尿病预测研究  

Diabetes prediction based on artificial intelligence

作　　者：周乐明 尚明生[2] 王永红 宋景麟 李小松 黄刚 王科 Zhou Leming;Shang Mingsheng;Wang Yonghong;Song Jinglin;Li Xiaosong;Huang Gang;Wang Ke(College of Computer Science and Technology,Chongqing University of Posts and Telecommunications;Big Data Center,Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences;Clinical Laboratory,Qianjiang Central Hospital of Chongqing;Bank of Chongqing Postdoctoral Research Center;Department of Cardiovascular Medicine,The Third People’s Hospital of Chengdu;Clinical Laboratory,Yongchuan People’s Hospital of Chongqing)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]中国科学院重庆绿色智能技术研究院大数据中心,重庆400714 [3]重庆市黔江中心医院检验科,重庆409099 [4]重庆银行博士后研究中心,重庆400024 [5]成都市第三人民医院心血管内科,重庆610031 [6]重庆市永川区人民医院检验科,重庆402160

出　　处：《重庆医科大学学报》2023年第12期1489-1492,共4页Journal of Chongqing Medical University

基　　金：重庆市科技局、重庆市卫生健康委联合科研资助项目(编号:2019ZDXM006)。

摘　　要：目的:以临床类指标建立基于极限梯度增强(extreme gradient boosting,XGBoost)、基于梯度提升树的分类器(light gra-dient boosting machine,LightGBM)、自适应增强(adaptive boosting,AdaBoost)、多层感知器(multilayer perceptron,MLP)等4种分类器的糖尿病预测模型,并评价其筛查效果。方法:根据病例对照研究设计采集研究组、对照组的99项临床类数据,使用py-thon3.8进行了分析,接着采用线性插补、固有非负隐特征(inherent non negative implicit features,INLF)模型等方法对特征缺失值进行了预测,然后使用4种分类器构建分类模型来检测糖尿病。结果:3241例高血压合并糖尿病患者作为研究组,4181例高血压患者作为对照组被纳入模型进行分析,包含99个特征,通过基于XGBoost、LightGBM、AdaBoost和MLP等4种分类器的糖尿病鉴别分类准确率分别为0.8949、0.8875、0.8620、0.8566。结论:本研究提出基于INLF预测的分类器模型框架的筛查效果较好,初步解决了通过机器学习来进行糖尿病早期筛查的问题,对临床诊断具有一定的实际意义,可作为一种简单、有效的糖尿病及其并发症筛查的方法。Objective:To establish a diabetes prediction model based on four classifiers of extreme gradient boosting(XGBoost),light gradient boosting machine(LightGBM),adaptive boosting(AdaBoost),and multilayer perceptron(MLP)according to clinical indicators,and to evaluate the screening effect.Methods:According to the case-control study design,99 attributes of clinical data from the study group and the control group were collected,and analyzed by python 3.8.Then the linear interpolation method and an inherent non-negative latent feature(INLF)model were used to predict the feature missing value,and the classification model was constructed using four classifiers to detect diabetes.Results:Through analyses of 3241 patients with hypertension combined with diabetes(study group)and 4181 patients with hypertension(control group)in the model,99 features were included.The accuracy rates of the diabetes classification model based on XGBoost,LightGBM,AdaBoost,and MLP classifiers were 0.8949,0.8875,0.8620,and 0.8566,respectively.Conclusion:Our proposed classifier model framework based on INLF prediction has a good screening effect,and preliminarily solves the problem of early diabetes screening through machine learning,which has certain practical significance for clinical diagnosis and can be used as a simple and effective screening method for diabetes and its complications.

关 键 词：不完备数据 糖尿病并发症 固有非负隐特征 分类器 

分 类 号：R587.1[医药卫生—内分泌]