Constructing Prediction Model of Lung Cancer Treatment Survival

According to statistics of the Ministry of Health and Welfare (Taiwan) in 2016, the first of Top Ten causes of death was cancer. In this category, percentages of trachea, bronchus and lung cancer were the highest. Thus, it is particularly important to study the disease. Coverage rate of national health insurance in Taiwan is up to 99.6%. Database of the national health insurance includes insurance information of the public. It is currently the mostly complete, mature and continuous clinical medial empirical data. With precise data and sufficient samples, they become the most appropriate and precise data for prediction of effectiveness of clinical illness treatment and comorbidity analysis. This study focuses on non-small cell lung cancer. It obtains cancer registration files from two million files in database of national health insurance, and screens patients who have suffered from ICD C33-34 “trachea, bronchus and lung cancer”. It then connects with OPDTE of outpatient services of national health insurance to include factors of comorbidity of disease in model design and treat DEATH as analytical figure after treatment and prognosis. By linear and non-linear data mining, it studies effects of different therapies on survival rate of lung cancer patients in three years and establishes prediction model. Through Artificial Neural Network (ANN) and Logistic Regression (LR), this study successfully establishes prediction model of survival rate of treatment of non-small cell lung cancer. When the best point of LR model: Sensitivity = 77.3 and Specificity = 73.8, it includes 81% (AUROC, Area Under ROC curve) of model precision rate. When the best point of ANN model: Sensitivity = 77.6 and Specificity = 76.8, it can include 83% (AUROC) model precision rate. It shows that model precision rate of ANN is higher than logistic regression model.