| Summary: | 碩士 === 國立虎尾科技大學 === 工業工程與管理研究所 === 101 === In Taiwan, retinal detachment is one of the main causes for people’s blindness. It is estimated approximate 1,000-2,000 patients suffering from such disease with the highest incidence among countries of the Far East. Relevant literature has indicated such phenomenon might be related to the rapid growth of myopia population. Causes of retinal detachment can be divided into three types: rhegmatogenous, traction and exudative retinal detachment, in which rhegmatogenous retinal detachment, commonly known as retinal detachment, must be treated with surgery as soon as possible. Thus, this study would focus on rhegmatogenous retinal detachment to carry out a further discussion. Patients’ retinal detachment may reoccur at the same location after receiving the first time surgery and the majority of them occur within one year (readmission rate at about 10 to 20%) which results in the need to accept another one or more surgeries. Therefore, effective evaluation of postoperative retinal detachment readmission will provide substantial assistance to relevant medical institutions, staffs, patients and their families as well as help medical resources allocation and control.
This research focused on readmission patients within one year after receiving RRD operation whose data were provided by 2007-2010 National Health Insurance Research Database as well as collected and compiled related impact factors through literature study. This study used Back Propagation Neural Network (BPN), Support Vector Machine (SVM), Particle Swarm Optimization (PSO), Genetic Algorithm-Logistic Regression, (GALR) and their mutual combination methods to construct seven groups of prediction models and applied Case Based Reasoning (CBR) to build evaluation system. Research findings showed that individual model in the paired samples T-test comparison presented significant differences, in which the prediction performance of Back Propagation Neural Network and Genetic Algorithm-Logistic Regression combining with Back Propagation Neural Network model was the most outstanding, at an average accuracy and Average Area under the Receiver Operative Characteristic Curve (AURC) more than 85% and 0.81, respectively. In the case based reasoning systems, the evaluation performance of introducing the weight value analyzed by GALR was the best, with an average similarity of 99.58%, system accuracy of 82.66% and 0.7372 of area under the curve of ROC. Therefore, results of this study should be able to be used as a reference for the prevention of postoperative readmission as well as to provide substantial improvement for the quality of postoperative care.
|
|---|
