Evaluate Glomerular Filtration Rate By The Radial Basis Function Neural Network In Patients With Chronic Kidney Disease

Objective: Accurate estimates of glomerular filtration rate?GFR?in patients with chronic kidney disease?CKD?are very important in clinical practice.A lot of methods of evaluation are available,but most of these methods can't reflect the GFR factually.For this reason,we developed a radial basis function?RBF?neural network and investigated the performance of this method in the estimation of GFR for patients with CKD.Methods: Selecting 651 patients with CKD who were hospitalized in nephrology department in June 2012 to February 2016 and had not accepted renal replacement therapy.All patients were given 99 m Tc-DTPA renal dynamic imaging examination,standard glomerular filtration rate?s GFR?was determined by ^99mTc-DTPA renal dynamic imaging method.Collected the materials including the patient's gender,age,blood urea nitrogen?BUN?,serum creatinine?Scr?,albumin?ALB?at the same time.The original data was processed by normalization and was divided into the training set?460 cases?and testing set?191 cases?in 2.4: 1 ratio randomly.Training 460 patients data taking s GFR as standard values to establish a RBF neural network which was used to estimate GFR,validating the model using 191 testing set data and estimating GFR to get eGFR_RBF.GFR was also estimated by the 4-variable Modification of Diet in Renal Disease?MDRD?equation and the 6-variable MDRD equation to get e GFR₄?e GFR₆ respectively.Comparing GFR among CKD stages,analysing the correlation between eGFR_RBF,e GFR₄,e GFR₆ and s GFR and the deviation of GFR which were estimated by three kinds of methods in difference,absolute difference,absolute deviation rate,15%?30%?50% coincidence rate,mean difference,95% limits of agreement,etc.Eventually,described the accuracy and precision of a variety of methods which were used to assess GFR and evaluated the performance of RBF neural network in the estimation of GFR for patients with CKD.Results:The correlation analysis between e GFR and s GFR showed that e GFR₄,e GFR₆,eGFR_RBF and s GFR had close correlation?r=0.894,0.910,0.912??P<0.05?.There was no statistical differences between eGFR_RBF values and s GFR values for stage 2-5 CKD?P>0.0125?.What's more,eGFR_RBF values were less than s GFR values for stage1 CKD?P <0.0125?and e GFR₄ values and e GFR₆ values were less than s GFR values for stage4 CKD?P <0.0125?,while the differences had no statistical mean between e GFR₄ values and e GFR₆ values and s GFR values for other CKD stages?P> 0.0125?.This result suggested that RBF neural network can evaluate GFR greatly for patients with 2-5 stages CKD,and it has no difference with 99 m Tc-DTPA renal dynamic imaging method,but it underestimates GFR for patients with 1 stages CKD.In addaiton,the 4-variable MDRD equation and the 6-variable MDRD equation can evaluate GFR greatly for patients with 1-3 stages CKD,but it underestimates GFR for patients with 4 stage CKD.In the field of evaluating the accuracy of GFR,the absolute difference median,the difference absolute median,absolute deviation rate which eGFR_RBF deviated s GFR between eGFR_RBF and s GFR was significantly less than that between e GFR₄ and s GFR,e GFR₆ and s GFR?P <0.0167?.Meanwhile,eGFR_RBF estimates within 15%,30%,and 50% for s GFR had high percentage and it had significant difference in stage 2,4 CKD especially.It showed that RBF neural network can provide more accurate GFR estimates than 4-variable MDRD equation and the 6-variable MDRD equation for patients with stage 2,4 CKDIn the field of evaluating the precision of GFR,the interquartile range of difference,the interquartile range of absolute difference,95% limits of agreement between eGFR_RBF and s GFR was less than that between e GFR₄ and s GFR,e GFR₆ and s GFR,especially in stage 1-3 CKD,showed better precision for the RBF neural network,especially provided more precise GFR estimates for patients with stage 1-3 CKD.In addation,for stage 5 CKD,the the absolute difference median,the difference absolute median,mean difference,absolute deviation rate which eGFR_RBF deviated s GFR between eGFR_RBF and s GFR was significantly more than that between e GFR₄ and s GFR,e GFR₆ and s GFR?P <0.0167?,what's more,the interquartile range of difference,the interquartile range of absolute difference,95% limits of agreement between eGFR_RBF and s GFR was more than that between e GFR₄ and s GFR,e GFR₆ and s GFR.Meanwhile,eGFR_RBF estimates within 15%,30%,and 50% for s GFR was less than that e GFR₄ and e GFR₆?P <0.0167?,but the difference had no statistical mean between e GFR₄ and s GFR,e GFR₆ and s GFR?P?0.0167?.This result suggested that the 4-variable MDRD equation and the 6-variable MDRD equation can provided more precise and more accurate GFR estimates than the RBF neural network for patients with stage 5 CKD.Conclusions:1.RBF neural network can evaluate GFR greatly for patients with 2-5 stages CKD,and it has no difference with 99 m Tc-DTPA renal dynamic imaging method,especially can provide more accurate and more p precise GFR estimates than 4-variable MDRD equation and the 6-variable MDRD equation for patients with CKD stage 2,but it underestimates GFR for patients with 1 stage CKD.2.The 4-variable MDRD equation and the 6-variable MDRD equation can evaluate GFR greatly for patients with CKD stage 1-3 and stage 5,especially can provide more accurate and more p precise GFR estimates than RBF neural network,but it underestimates GFR for patients with 4 stage CKD.