Application Of Improved RBF Neural Networks On Glomerular Filtration Rate Estimation

Artificial neural network is a technique of modeling which has the ability to approximate nonlinear functions and learn from experiences. It shows great advantages for regression prob-lems. The Radial Basis Function(RBF) networks are a type of neural network which is easier to be designed and trained. They have good generalization properties and high tolerance to in-put noise, which enhances the stability of the model. Therefore, RBF neural network has been applied to many fields in recent years, such as image precessing, fault diagnosis and pattern recognition.Considering the great properties of RBF neural network, this paper presents a hybrid al-gorithm. On the one hand, the output weights, the centers and widths of RBF as well as input weights are adjusted during the training process by an improved Levenberg-Marquardt algo-rithm. On the other hand, in order to obtain the best training result, it is a key process to find good initialization of RBF units and proper size of architecture. Therefore, an incremental design scheme is proposed to find a compact network, which shows good stability and generalization.Accurate estimation of Glomerular Filtration Rate(GFR) in patients with Chronic Kidney Disease(CKD) is a key process in clinical practice. In cooperation with the nephrology depart-ment of the Second Hospital of Dalian Medical University, an RBF neural network is build to estimate GFR on the basis of data collection and related medical research. The improved RBF network achieved better performance compared with other models and traditional equations.The proposed GFR estimation model is more accurate and reliable than traditional equa-tions, and has high practical application value in the prevention and cure of CKD. At the same time the approach of neural network in medical applications is superior to the traditional statis-tical analysis and provides a new way for GFR estimation.