Research Of Estimating Glomerular Filtration Rate Based On Artificial Neural Network

Chronic kidney disease (CKD) is a serious disease with high morbidity. Glomerularfiltration rate (GFR) is used clinically to evaluate renal function during assistant diagnosis.99mTc-DTPA is considered as a gold standard for testing GFR because of high accuracy, but itsclinical promotion is limited because expensive equipment is absolutely necessary. In order tofind a convenient, quick, cheap way to access GFR, which could also be implemented ingeneral hospital, empirical equations for estimating GFR were presented by Americanscholars, and had already applied widely in America. However, the performance of theseequations was poor in estimating Chinese people’s GFR and was far less than the clinicalrequirement. The progression of domestic related research was slow. This paper is based onthe cooperation with the nephrology department of the Third Affiliated Hospital of SunYat-Sen University. On the basis of data collected in follow-up visit and related research,methods of artificial neural network are applied to estimate GFR.A total of1180samples are divided into training data set, internal and external validationdata set. After pretreatment, BP network, GRNN, Legendre neural network, and polynomialneural network are applied to estimate GFR, respectively. BP network is optimized by geneticalgorithm, and S function is used in Legendre neural network to extend the input variables.An adaptive algorithm is presented to determine the combination of power of input variablesin polynomial neural network. Then MIV analysis is applied in BP network optimized bygenetic algorithm for selecting input variables in order to improve the performance.The performance is evaluated mainly from aspect of accuracy and consistency. Accuracyis represented by deviation, percentage of deviation, and agreement rate, and consistency isrepresented by Bland-Altman Plot. Statistically significant of difference is examined by thecorresponding statistics examination methods. Finally the optimized BP network with6inputvariables is chosen because of its better performance in the comparison and a graphical userinterface is designed to display the network.The proposed network is superior to the traditional equations, and could be applied inclinical situation for further verification and update. At the same time the result shows that themethod of machine learning represented by artificial neural network is superior to classicalstatistical method, and has advantage in biomedical data processing.