Constructing the Models for Chronic Kidney Disease Diagnosis and Risk Estimation by Applying Neural Network Technologies

• Main Authors: Yu-Chin Chen, 陳宥瑾
• Other Authors: 邱瑞科
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/23123092319346022106

碩士 === 輔仁大學 === 資訊管理學系 === 98 === According to the 2009 Annual Data Report of United States Renal Data System, the report indicated that the rates of incidence and prevalence in end stage renal disease (ESRD) in Taiwan are both ranked in the first place in the world. The symptoms of chronic kidney disease (CKD) are not apparent, it is difficult to be sensed under general condition. It is probably the bad habit of daily life resulting in gradual declination of kidney. It probably is ESRD once it is found, it is required to rely on the dialysis to maintain alive. Therefore, the early diagnosis and prevention of CKD is extremely important. By reviewing the literatures and expert interviews, the diagnostic factors of causing CKD used in the computational formula which the physician applies and the potential impact factors of CKD are identified. Based on these factors, the back-propagation network (BPN), generalized feedforward neural networks (GRNN), modular neural network (MNN) are employed to determine whether the people affect with CKD. Then, the three types of neural networks respectively combined with genetic algorithm (GA) are furtherly employed. In addition to diagnose whether the people affect with CKD. It is used to observe whether the models can reach better performance in the identification of CKD. Lastly, the fuzzy expert system is employed to develop a support system for estimating risk of affect with CKD. The results of the simulation of experiment and assessment of performance show that the diagnostic factors and the potential impact factors are the two that have the best performance in BPN showing the 100% accuracy for classification. The diagnostic factors have the best performance by combining BPN and GA that have a 91.71% of accuracy for classification. While the potential impact factors have the best performance by combining GRNN and GA that have a 78.39% of accuracy for classification. It is prove that the potential impact factors such as creatinine, blood glucose, blood pressure, blood urea nitrogen (BUN), proteinuria and hematuria could influence CKD strongly through the experiments of neural networks. Estimation for the risk of causing CKD could be inferred by applying this system, it can also equip medical staff with the ability to make precise diagnosis so as to let public clearly understand the risk level of causing CKD in order to take necessary precaution in advance to avoid any risk of CKD or prevent their condition from getting worse which is also what the philosophy of preventive medicine is all about.