決策樹於中西醫腦中風診斷指標結合之應用

An Application of Decision Tree Methodology to Stroke Diagnoses Using Combined Traditional Chinese and Modern Medicine Indices

背景與目的－－在過去的二十年裏，中風一直都是國人三大死亡原因之一，也是造成中老年人殘障的最主要原因。大部分的中風患者常發生神經障礙後遺症，對於個人、家庭、社會的影響相當鉅大。西醫對中風的治療主要著重於防止繼發性的腦細胞傷害及其併發症，而中醫主則對於腦中風的恢復期、後遺症期的康復治療累積了許多臨床經驗。因此結合中西醫在腦中風的康復治療方面之所長，進行科學化的研究分析，應是治療中風的良方。然而，有關腦中風中醫證候與西醫診察指標的相關性研究報告卻不是很多。因此，本研究欲建立西醫之腦中風診斷指標，與中醫的診斷指標之相關性，以提供腦中風患者更好的治療。方法－－本研究以資料探勘中的C4.5決策樹演算法與貝氏分類法來建構分類，把西醫在腦中風診斷的診察指標和中醫公認的診斷指標建立關聯。這些萃取出來的規則可以作為中醫師未來臨床診斷的參考依據。此外，我們也利用腦中風資料來探討因為訓練資料組類別分布不均勻與分類錯誤成本在C4.5分類器及貝氏分類法效能的影響。結果──在中西醫指標之連結方面，共萃取出八條規則。而在建立有效分類模型方面，以火證之研究結果發現，當C4.5分類器之訓練資料組的火熱證跟無火熱證的比例相同並與錯誤成本概念結合時，可建構出本實驗中具有最佳效能的分類模型。

The Department of Health recently published data indicating that stroke (Cerebral Vascular Accident, CVA) ranked third among the main causes of death in Taiwan with a mortality rate of 9.23 per 100,000 in 2004. It is also the main cause of disability among the elderly since varying degrees of neural sequela usually remain, leaving a lasting impact on the individuals, family and society. Though the Traditional Chinese Medicine (TCM) has established a stroke diagnosis standard supported by the TCM dialectical system, very few studies were able to establish referential relationships between these standards and those used by modern medicine. In this research, we have analyzed the relationship between diagnostic indices used in TCM and modern medicine and apply the C4.5 decision tree classifier (algorithm) and Bayesian classifier to extract rules used in stroke diagnoses. Furthermore, we look into the proportion of fire syndrome to non-fire syndrome in the training set to determine the misclassification cost that may influence the performance of the classifier. Our analysis shows that when the number of fire syndrome is proportioned to the number of non-fire syndrome in the training set, the classifier has better results. And when we combine the class distribution and misclassification cost, the classifier may improve sensitivity without reducing its specificity.