[[alternative]]Using Non-invasive Data to Implement an Artificial Neural Network Diabetes Computer Aided Diagnosis System

[[abstract]]根據美國疾病管制與預防中心於2017年所發佈的糖尿病統計年報指出糖尿病人口在2012年，平均每年個人醫療花費約為1.37萬美元，其中有7900美元是由糖尿病所造成。根據2002年美國糖尿病協會於糖尿病照護期刊指出提早發現糖尿病，能夠減少醫療與患者的負擔。現今量測糖尿病的方式是利用侵入式的血液取樣，但其測量方式需要測量工具與使用、判讀知識才能進行測試，更有學者指出，低教育與低收入的美國民眾，不知何謂血液取樣測試。另外，雖然目前機器學習應用於糖尿病輔助預測有相當高的預測準確度，但其模型多數仍使用侵入式檢測後的生理數據做為訓練資料。本研究提出一個利用隸屬美國疾病管制中心的（Center of Disease and Prevention,CDC）危險因子監控系統（Behavioral Risk Factor Surveillance System）之非侵入式資料來建立人工類神經網路，並依照不同的訓練資料與隱藏層數，實作出50個模型，最後依照模型效度選用了11個模型做為糖尿病的輔助預測模型，並以模型為基礎建置糖尿病線上電腦輔助診斷系統，提供問卷給予使用者填答，並能夠將使用者輸入的答案利用模型進行分類，如模型分類結果為患有糖尿病風險，則系統將會顯示提醒訊息告知使用者盡速就醫檢查，本研究成果能夠提供民眾在無法使用侵入式血液取樣下，提供是否罹患糖尿病的預測，藉此輔助民眾是否進行就醫的判斷，早日取得治療契機，提升治療效果，節省國家醫療資源。[[abstract]]According to the National Diabetes Statistics Report released by the Centers for Disease Control and Prevention of the USA in 2017, the average medical cost of each person is 13.7 thousand dollars which 7.9 thousand was paid for diabetes-related treatment. In 2002, the American diabetes association indicated that early detection of diabetes can reduce the medical expense and health cost for diabetes patient as well. Nowadays, the invasive blood test is the main method to do diabetes assessment, which needed the measurement tools, knowledge of tool operation and result from discrimination. Furthermore, many reports indicate that low income and less educated people in the United States do not have the sense of blood test. Bseides, when machine learning is applied to diabetes prediction, most of the model still employ invasive blood sampling as analyzing data for model generation. This study employed non-invasive data from Behavioral Risk Factor Surveillance System to establish a set of artificial neural network aided of diabetes prediction models. According to different training data and hidden layers, 50 models are made and , 11 models are selected as prediction model after model accuracy evaluation. Based on these models, this study implements an online computer aided diagnosis system which provides a self-assessment questionnaire to usres and submit the answers to the model for classification, if the classification result shows a high risk of diabetes, the system will display a reminding message suggest the user to seek for professional medical examination as soon as possible. This study results can provide people with a set of model to aid the prediction of diabetes without invasive blood sampling, and to enable early detection of diabetes patients for immediate treatment and recovering. Moreover, it can promote government’s medical resources effectiveness