WASP neuronet activated by bipolar-sigmoid functions and applied to glomerular-filtration-rate estimation

By combining two fast training methods, i.e, the weights-direct-determination (WDD) method and Levenberg-Marquardt method, this paper proposes a novel training algorithm called weights and structure policy (WASP) for the three-layer feedforward neuronet, in addition to the algorithm of weights and structure determination (WASD). Note that the pruning-while-growing and second-pruning techniques are developed and exploited in the WASP algorithm with the aim of achieving a neuronet with a simple and economical structure. In order to verify the WASP efficacy and to address the problem of chronic kidney disease (CKD) for clinical applications in China, numerical experiments about estimating glomerular filtration rate (GFR) by the WASP neuronet and traditional GFR-estimation equations are conducted and compared. The experiment results show that the WASP training speed is fast and that the estimating accuracy via the WASP neuronet is around 20% higher than those via traditional GFR-estimation equations. The WASP efficacy is thus demonstrated with a significant value in GFR estimation of CKD for clinical applications.