本論文的目的是透過不同機器學習的方式，包含單純貝氏分類法、多層感知器、以決策樹為分類器的引導聚集算法、支持向量機等演算法，來建立臨床失智症量表數據和醫師診斷的篩選的模型。臨床量表數據依據失智症的嚴重程度分成六類、五類和三類等三種，研究結果顯示以決策樹為分類器的引導聚集算法為最佳。由於三類數據分布明顯為非平衡，因此透過SMOTE方法對，對少量數據進行調整，分析結果顯示對準確度與精確率的提升相當有限。最後利用主成分分析的技術，進行臨床量表題目的幾種化簡，並經由信度分析，選擇臨床量表的重要題目，研究結果顯示語言有關題目扮演重要角色。 The purpose of thesis is to establish CDR score data and screening models for physician diagnosis through different machine learning methods, including classifiers such as Naïve Bayes, Multilayer perceptron, Bootstrap aggregating with decision tree, and support vector machine. The CDR score data are divided into six, five and three stages according to the severity of dementia. Bootstrap aggregating with decision tree classifier is the best among the others. Also, the population sizes for three stages case are obviously imbalanced, SMOTE method is used to adjust a small amount for the normal statege and the study result shows the improvement is rather limited. Finally, principal component analysis (PCA) is carried to to simplify CDR questionares, and certain questions are selected and verified through the reliability analysis. Our study shows that newly added language-assesment questionnaire plays an important role in our analysis.
