基于体检数据机器学习分析的糖尿病风险预测模型

Diabetes Risk Prediction Model Based on Machine Learning Analysis and Physical Examination Data

针对糖尿病风险预测中数据单一导致预测误差较大的问题,本研究基于体检电子病历数据分析搭建空腹血糖预测模型,探究适合进行空腹血糖预测建模的方法,预测血糖指标及糖尿病的患病风险。基于数据挖掘基本流程,进行数据预处理,采用序列后向算法进行特征选择,使用决策树、随机森林、SVM、逻辑回归及朴素贝叶斯分类5种机器学习算法进行建模预测,并验证所构建模型的效果。研究结果表明,五种算法的准确率均高于88%,其中SVM准确率最高,达96.7%;敏感度均高于66%,随机森林敏感度最高,为95.1%;特异度均高于88%,逻辑回归特异度最高,为97.0%;AUC的值均高于0.8,随机森林最高为0.942。综合比较序列后向选择算法,随机森林算法更适合搭建糖尿病风险预测模型。该研究对通过电子病历数据进行空腹血糖预测的准确度更高,具有很高的应用价值。

Aiming at the problem of large prediction error caused by single data in diabetes risk prediction,this research builds a prediction model of fasting blood glucose by analyzing the data of electronic medical records,explores a method suitable for fasting blood glucose prediction modeling based on electronic medical records data,and predicts blood glucose index and the risk of diabetes.Based on the basic process of data mining,data preprocessing is carried out.Sequence backward selection is used for feature selection.Five machine learning algorithms,including decision tree,random forest,SVM,logistic regression and naive Bayes classification,are used for modeling and prediction,and the effect of the constructed model is verified.The results show that,the accuracy rates of the five algorithms are all higher than 88%,of which the highest is 96.7%for support vector machine.The sensitivity is higher than 66%,of which the highest for is 95.1%for random forest,and the support vector machine lowest is 66.4%.The specificity is all higher than 88%,of which the highest is 97.0%for logistic regression.The area under the ROC curve is all higher than 0.8,of which the highest is 0.942 for random forest.Through comprehensive comparison using SBS feature selection algorithm,random forest algorithm is more suitable for building diabetes risk prediction model.This research has a higher accuracy in predicting fasting blood glucose through electronic medical record data,and has a high application value.