運用基因演算法建構糖尿病前期診斷模型之研究

The Use of Genetic Algorithm to Predict Prediabetic Disease

在國人十大死因中，腦血管疾病、心臟血管、肝臟病變、腎臟病變、高血壓等慢性疾病佔有極高比例，而糖尿病患者合併上述疾病的發生率比一般人高出甚多。雖然慢性疾病未必立即造成死亡，但其逐漸侵蝕中年與老年人之健康，使其成為高就醫頻率的族群，不但耗費更多的醫療資源，也將嚴重影響其生活品質。就預防醫學的角度而言，建立一套可以準確診斷出國人由正常血糖發展為糖尿病前期 (Prediabetes)的疾病預測模型有其必要性。但是，現今疾病診斷或預測技術，對於疾病影響變因之複雜度與變異性無法有效的掌握，而演化式計算中的基因演算法(Genetic Algorithms, GA)適合用來解決複雜、NP-hard的問題。本研究將以病患就醫過程中產生的相關資料為基礎，建立一可以早期診斷出糖尿病前期的診斷系統，提供醫師對病患診治之參考，對不同疾病的程度提供適當的治療與建議，延緩或預防其併發症的進行，進而達到預防性醫學的目標。從研究成果顯示，GA模型、NN模型、C 5.0模型等屬人工智慧領域技術的表現，明顯優於普遍被流行病學者用來建立疾病預測模型的Logistic Regression方法；而運用GA建立的模型，其診斷預測績效比Neural Network、C 5.0技術所建立的模型好。

Hypertension, heart diseases, cerebrovascular accident and chronic renal insufficiency are ranked among the leading 10 causes of death of Taiwanese in recent two decades. Patient with diabetes mellitus usually have higher probability to contract the above-mentioned disorders than normal populations. Although in most instances, chronic diseases do not cause death immediately, they will progressively do harm to the health and quality of life of middle-aged people and elderly. The patients with diabetes mellitus and other chronic diseases would form a high hospital-visiting population and consume a lot of medical resources. From the point of view of preventive medicine, establishing a standard model to predict the possibility of development from euglycemia to prediabetic state is necessary, however, the current technology can't efficiently handle the complexity and variability of the factors affecting the natural progress of the disease. We know that in calculation methods, Genetic Algorithm (GA) is appropriate to resolve complicated, NP-hard problems. In this study, we tried to establish a diagnostic system for early diagnosis of prediabetes by means of computerized analyzing the laboratory data of the patients during medical visits. We found that for disease prediction, the models using artificial intelligence techniques, including GA, NN and C 5.0, are obviously better than Logistic Regression method that is usually applied in conventional epidemiology. In addition, the GA model is more sensitive than NN and C 5.0 models. Nevertheless, we hoped that these models could provide useful information to help physicians in disease prevention so as to delay or diminish occurrence of complications or sequels.