應用人工智慧與邏輯斯迴歸於胰臟癌輔助鑑別診斷之比較分析

A Comparative Analysis of Artificial Intelligence and Logistic Regression for Assistance in Differential Diagnostic of Pancreatic Cancer

徐銘遠(MING-YUAN HSU)

國立虎尾科技大學/管理學院/工業工程與管理研究所/碩士(2006年)

摘要

近年來醫學研究指出，胰臟癌占國人十大癌症死因的第八位，台灣每年因此癌死亡人數亦達 800人，且有逐年增加的趨勢。因為大部分胰臟癌早期之症狀多不具特異性，醫師的人為判斷因每個人的經驗和認知有所不同及當時的精神狀況影響，難免有診斷上的誤差而導致了後續的檢驗變的沒有幫助，不僅造成了醫療資源的浪費更嚴重的延誤了病患的病情，錯失了「早期診斷，早期治療」的時間。本研究以人工智慧方法中的類神經網路與基因演算法及統計方法中的邏輯斯迴歸建構三種鑑別模式，用以鑑別胰臟癌與急性胰臟炎。並以ROC曲線對此三種模式的鑑別能力作比較與分析。本研究使用234筆病例資料作為訓練樣本另以117筆病例資料作為測試資料，從成對比較分析的結果顯示，三種模式之AUC值，除了GALR模式有明顯優於SLR模式之外，無論是SLR模式與BPN模式的成對比較或BPN模式與GALR模式的成對比較結果，均無明顯的差異，但若在三種模式的最佳門檻值的要求之下，GALR模式之敏感度和特異度為96.7%和82.5%優於SLR模式之敏感度和特異度為96.7%和73.7%優於BPN模式之敏感度和特異度為88.3%和84.2%，結果顯示，GALR模式的表現結果較佳。最後，未來如果有更龐大更完整之資料庫及更精確的運算方法，人工智慧方法將會有更佳的預測效果。

並列摘要

In Recent medical report, pancreas cancer has reached the 8th in leading cause of cancer death in Taiwan. Based on the epidemiology study, there is a high incidence of the pancreatic cancer in Euramerica, almost one out of ten thousands in morbidity. About 800 people die of this cancer each year in Taiwan, and the figure is up trending. Most of the prognosis for pancreas cancer is not specificity, such as in appetence, abdominal distension or pains are some demonstrations of characteristic clinical signs. Many clinicians treat patients with gastro duodenal or gallstone disease. It was not until the increase in pains or the pains extended to the dorsal that the tumor was found to invade the nearby organ, result in the lost of treatment in time.
When dealing with the abdominal disease, clinicians have to infer the related organ that causes the symptoms, physical examination and lab test were further taken. However, the characteristic of non specificity of the prognosis for pancreas cancer, clinical judgments vary with individual experience and perception or mental conditions usually causes the diagnosis bias.
This research applied Neural Network and Genetic Algorithm of Artificial Intelligence and Logistic Regression in Statistics to build three differential diagnostic models of Pancreatic Cancer and Acute Pancreatitis. The performance of all models was compared with Receiver Operating Characteristic (ROC) analysis. There are 234 cases for model trained and 117 cases for further testing. The results of the pairwise comparison shown that GALR model and SLR model had significant difference. However, the sensitivity and specificity of the GALR model is 96.7% and 82.5% which is better than SLR model of 96.7% and 73.7%, also BPN model of 88.3% and 84.2% at the optimal threshold for the three models. Which shows that GALR model has the better diagnostic performance. In the meantime, Artificial Intelligence will perform more accurate prediction when in larger and completed database along with advanced computing techniques.

參考文獻

3.陳漢明(1997)，胰臟癌之最新發展（Ⅲ）臨床表現、診斷及分期，當代醫學，二十四卷，第五期，17-21頁

6.章明珠、林肇堂(2002)，胰臟癌診斷的最新進展，當代醫學、二十九卷、第六期，445-448頁

7.陳建華(2003)，膽道、胰臟內視鏡鏡超音波術，健康世界月刊，326期，31頁

14.李文修(2005)，運用基因演算法建構糖尿病前期診斷模型之研究，醫療資訊雜誌，第14卷，第1期，頁51-64

26.陳建勝、林明宏(2003)，壽險保單早期失效之預測，風險管理學報，第五卷，第三期，pp.341-361