機器學習在血糖指數建模和預測中的應用

Abstract

本研究的目的是利用最小絕對值收斂和選擇算子（LASSO）與典型相關分析（CCA）來篩選餐後血糖（GLU_2hr_PC）、空腹血糖（GLU_AC）和糖化血紅蛋白（HbA1c）這三種血糖指數之潛在因子。同時，應用了羅吉斯回歸（LR）、K-近鄰演算法（KNN）、決策樹（DT）、線性判別分析（LDA）和支持向量機（SVM）五種不同的機器學習方法，對健檢資料集的患者是否有糖尿病進行分類。然後，將混淆矩陣建立的相關指標用於比較這五項分類器的表現。從研究結果可以發現，糖尿病的危險因素可能包括鈣離子（Ca）、氯離子（Cl）、肌酸酐 （CRE）、高敏感度C-反應蛋白（hs_CRP）、尿比重（Sp_Gr_C）、年齡、身高體重指數…等等。而在機器學習的分類結果裡，綜觀所有的評分指標，LDA在此資料集對於糖尿病病患的偵測能力最佳。這項研究將提高我們對於健康檢查的重視，並且針對降低糖尿病病例的臨床決策有所幫助。

The objective of this study is to identify the potential factors of glycemic indices, i.e., postprandial blood glucose (GLU_2hr_PC), fasting blood glucose (GLU_AC), and glycated hemoglobin (HbA1c) by least absolute shrinkage and selection operator (LASSO) and canonical correlation analysis (CCA). At the same time, five different machine learning methods including logistic regression (LR), K-nearest neighbor classification (KNN), decision tree (DT), linear discriminant analysis (LDA), and support vector machine (SVM) were applied to build the learning system in order to classify whether the patients in the healthcare dataset have diabetes or not. Then, the relevant indicators established by the confusion matrix were used to compare the performances of these five classifiers. It can be found from the research results that the risk factors for diabetes may include calcium (Ca), chloride (Cl), creatinine(CRE), high sensitivity C-reactive protein（hs_CRP）, urine Specific Gravity（Sp_Gr_C）, age, Body Mass Index, etc. In the classification results of machine learning, LDA has the best detection ability for diabetic patients in this dataset among all the evaluate indicators. This work will increase people’s attention to the health checkups and aid in the clinical decision making for reducing diabetes cases.