應用類神經網路預測鼻咽癌病人之五年存活狀態

Clinical Application of Artificial Neural Network in Predicting Five-Year Survival of Patients with Nasopharyngeal Carcinoma

鄭中浩(Chung-Hao Cheng)

指導教授 ： 邱泓文

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2013年)

https://doi.org/10.6831/TMU.2013.00039

摘要

鼻咽癌為好發於中國東南沿海的惡性腫瘤，臨床上需要準確的鼻咽癌預後評估系統，提供個人化的存活預測資訊，以幫助醫師和病人共同決定如何面對及治療鼻咽癌。本研究篩選1990年至2005年間於臺灣地區某癌症專科醫院確診的1,114位鼻咽癌病人，選用年齡、性別、鼻咽癌癌症期別、腫瘤侵犯程度、淋巴結擴散狀態、遠端轉移狀態、病理細胞切片分類、放射治療方式、化學治療方式、乳酸脫氫脢指數、鹼性磷酸脢指數、過往吸菸習慣、家族鼻咽癌病史等十三項變數，剔除變數資料欄位不全者70人後，分為包含遠端轉移（1,044人）及排除遠端轉移（984人）的兩組研究樣本，分別建立類神經網路模型以預測鼻咽癌病人的五年存活狀態，隨機挑選75%為訓練組樣本，25%為測試組樣本，並對其預測結果準確率、靈敏度、特異度和接收者操作特徵曲線下面積進行評估與分析。研究樣本平均年齡為45歲，兩組研究樣本的五年整體存活率分別為73.47%及77.74%，應用STATISTICA 10.0軟體以多層次感知器類神經網路方式建立起最佳化的預測模型分別為MLP 36-17-2及MLP 34-5-2，與真實病人存活狀態資料進行比對分析，兩組研究樣本中對於個別病人的存活狀態預測準確率分別為89.94%及90.96%，靈敏度為93.22%及93.73%，特異度為80.87%及81.28%，接收者操作特徵曲線下面積皆為0.95；而測試組樣本的預測準確率分別為90.04%及87.80%，靈敏度為94.85%及92.23%，特異度為76.12%及71.70%，接收者操作特徵曲線下面積為0.91及0.88。本研究應用類神經網路提供一種預測個別鼻咽癌病人五年存活狀態的方法，並發現選用較多輸入變項資料的類神經網路模型，預測表現較先前相關研究結果或其他統計工具更為良好，顯示類神經網路為預測個別鼻咽癌病人五年存活狀態的良好決策支援工具，可以對於鼻咽癌病人提供個人化的存活狀況預測資訊。

關鍵字

存活預測 ； 鼻咽癌 ； 類神經網路

並列摘要

Nasopharyngeal carcinoma is a malignancy with a high incidence in Southeast China. There is a need of survival predictions at the individual level to help doctors and patients make informed decisions together. In this research, 1,114 patients with nasopharyngeal carcinoma in one cancer center during the year from 1990 to 2005 are enrolled. The chosen variables include age, sex, clinical stage, primary tumor, regional lymph nodes, distant metastasis, biopsy, radiotherapy, chemotherapy, lactic dehydrogenase, alkaline phosphatase, smoking, and family history. The final samples are 1,044 patients excluding 70 patients with any data columns missing. Two artificial neural networks are created by computer software to predict the five-year survival status of those nasopharyngeal carcinoma patients with or without distant metastasis. Seventy-five percent of these patients are randomly selected and classified to training groups. The performances of prediction models are evaluated according to parameters such as accuracy, sensitivity, specificity, and the area under receiver operating characteristic curve. The average age of these patients is 45 years old, and the five-year overall survival rates are 73.47% and 77.74%. The optimized artificial neural networks are MLP 36-17-2 and MLP 34-5-2. Their accuracy is 89.94% and 90.96%, sensitivity is 93.22% and 93.73%, specificity is 80.87% and 81.28, and areas under the receiver operating characteristic curve are both 0.95 of those two groups. And their accuracy is 90.04% and 87.80%, sensitivity is 94.85% and 92.23%, specificity is 76.12% and 71.70, and areas under the receiver operating characteristic curve are 0.91 and 0.88 of those two test groups. This research shows that prognostic prediction systems established by artificial neural networks have the potential to predict the five-year survival status of individual patients with nasopharyngeal carcinoma. With more input data, the performance of prediction models is better than previous researches. So this prognostic prediction system can provide useful information to help individual patients with nasopharyngeal carcinoma make informed decisions.

並列關鍵字

Artificial neural networks ； Nasopharyngeal carcinoma ； Survival prediction

參考文獻

羅華強（2001）。類神經網路─MATLAB的應用。新竹市：清蔚科技股份有限公司。

民國101年死因統計年報（2013年版）【資料檔】。臺北市：行政院衛生署。

行政院衛生署國民健康局癌症登記線上互動查詢系統（2013年版）【資料檔】。臺北市：行政院衛生署國民健康局。

Astion, M.L., Wener, M.H., Thomas, R.G., Hunder, G.G., Bloch, D.A. (1994). Application of neural networks to the classification of giant cell arteritis. Arthritis and Rheumatism, 37(5), 760-770.

Baxt, W. G. (1995). Application of artificial neural networks to clinical medicine. Lancet, 346(8983), 1135-1138.