以類神經網路建構再發性腦中風之預測模式

Predicting Recurrent Stroke via ANN Model

腦中風（stroke）為最常見威脅性命的神經疾病，也是造成全球死亡率增加的主要因素之一，全球五十億人口中約有兩千五百萬到三千萬人口曾經罹患腦中風，以死亡率而言，美國每年約有十六萬人因腦中風而死亡，而台灣每年約有一萬人因腦中風死亡。暫時性腦缺血的病患再度腦中風的機率為一般人的十倍，其死亡率更高達四分之一。此外，腦中風的復發也會明顯增加死亡率、無法工作人口比率與平均住院日。因此，本研究以典型相關分析瞭解腦中風危險因子與再發性腦中風嚴重度指標之相關性，並建構一套預測模式，作為醫師診療時參考的依據。由於預測腦中風再發生的過程是複雜多變的，變數間呈非 線性關係，所以本研究採用類神經網路技術，以 331 筆病歷資料進行資料分析，再將資料進行 3 摺交叉 驗證（3-fold cross validation），並比較羅吉斯迴歸與類神經網路之預測績效。分析結果發現，良好的血壓控制與腦中風復發時的嚴重程度有關連性，而且有心臟病史的病患會影響其預後的狀況。最後，本研究 以準確度、敏感度與特異度評估預測模式的績效，分析結果分別如下類神經網路模型之準確度 94.8％（敏 感度 93.9％，特異度 95.7％）；羅吉斯迴歸模型為 94.6％（敏感度 92.3％，特異度 96.3％）。

Stroke is one of the life-threatening neuropathy. It is the key factor to increase mortality rate over the world. Two thousands five hundred million people suffered from stroke on earth. Among these figures, one hundred sixty thousand people died of stroke in America and ten thousand people died each year in Taiwan. Patients who had Transient ischaemic attack(TIA) suffered from recurrent stroke is ten times more. Their mortality rate is higher than one-fourth. Additionally, stroke recurrence is a significant concern with regard to an increase in mortality, disability, and length of hospital stay. Thus, correlation analysis can be used to look at the relationships between stroke risk factors and the severity score of recurrent stroke. This built system assist the physicians in diagnosis. To predict the stroke recurrence is a complex task and it is a nonlinear relationship among many variables. We developed an ANN model to assist the physicians to predict the possibility of stroke recurrence. The study is retrospective by using information from a database of medical inpatients. Three hundred and thirty one patients’ records were used as sample. To achieve optimum performance, we use a three-fold cross validation procedure. Furthermore, we compared the performance of ANN against the logistic regression approach on the same dataset. Our results show that patients with well control blood pressure will have lower severity score. Finally, we evaluated the performance of models according to prediction accuracy, sensitivity and specificity.