Forecasting Diabetic Patients’ Churn – by Using Data Mining Techniques

• Main Authors: Meng-Fu Tsai, 蔡孟甫
• Other Authors: Chin-Shien Lin
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/e64wm9

碩士 === 國立中興大學 === 企業管理學系所 === 106 === In this study, data were came from the longitudinal inpatient and outpatient claims of 100 million sampled registry from Taiwan National Health Insurance Research Database from 2000 to 2004. The purpose of the study was to explore the construction of a model for the diabetes patients’ churn by using hierarchical linear model. The factors that the patient considers when choosing a hospital were classified into three levels: patient, doctor, and hospital to explain the effects of different levels of factors on patients’ churn. In addition, using data mining techniques to predict diabetic patients’ churn and find out the feature of patients’ churn as a reference for the medical institution managers to develop a customer retention strategy.Finally, the models were evaluated with indicators : Accuracy, Sensitivity, and pecificity. The empirical results show that the younger diabetic patients are, the more diabetic patients will churn, diabetic patients are in a mild condition are more likely to churn than diabetic patients are in a severe condition,diabetic patients pay less for medical expenses are more likely to churn than for high medical expenses, the longer Distance between hospital and diabetic patients’ home is, the more diabetic patients will churn, the worse doctor’s ability is, the more diabetic patients will churn, the worse medical reputation is, the more diabetic patients will churn, patients visits low hospital grade are more likely to churn than patients visits high hospital grade. The empirical results not only presents a more precise variable relationship by using hierarchical linear model, but also has important management implications for hospital’s customer relationship management in practice. The accuracy of data mining is ranked from high to low as 64.18% for neural networks, 62.79% for decision trees, and 62.4% for Logistic regression.