A Back Propagation Neural Network Approach to Estimate the Glomerular Filtration Rate in an Older Population

Background: The use of creatinine-based glomerular filtration rate (GFR)-estimating equations to evaluate kidney function in elderly individuals does not appear to offer any performance advantages. We therefore aimed to develop an accurate GFR-estimating tool for this age group.

Methods: Adults aged ≥ 65 years who underwent GFR measurement by technetium-99 m-diethylene triamine pentaacetic acid (Tc-DTPA) renal dynamic imaging were included. Data were randomly split into a training set containing 80% of the participants and a test set containing the remaining 20% of the subjects. The Back propagation neural network (BPNN) approach was used to derive a novel GFR estimation tool; then we compared the performance of the BPNN tool with six creatinine-based equations (Chronic Kidney Disease-Epidemiology Collaboration [CKD-EPI], European Kidney Function Consortium [EKFC], Berlin Initiative Study-1 [BIS1], Lund-Malmö Revised [LMR], Asian modified CKD-EPI, and Modification of Diet in Renal Disease [MDRD]) in the test cohort. Three equation performance criteria were considered: bias (difference between measured GFR and estimated GFR), precision (interquartile range [IQR] of the median difference), and accuracy P30 (percentage of GFR estimates that are within 30% of measured GFR).

Results: The study included 1,222 older adults. The mean age of both the training cohort (n = 978) and the test cohort (n = 244) was 72 ± 6 years, with 544 (55.6%) and 129 (52.9%) males, respectively. The median bias of BPNN was 2.06 ml/min/1.73 m, which was smaller than that of LMR (4.59 ml/min/1.73 m; p = 0.03), and higher than that of the Asian modified CKD-EPI (-1.43 ml/min/1.73 m; p = 0.02). The median bias between BPNN and each of CKD-EPI (2.19 ml/min/1.73 m; p = 0.31), EKFC (-1.41 ml/min/1.73 m; p = 0.26), BIS1 (0.64 ml/min/1.73 m; p = 0.99), and MDRD (1.11 ml/min/1.73 m; p = 0.45) was not significant. However, the BPNN had the highest precision IQR (14.31 ml/min/1.73 m) and the greatest accuracy P30 among all equations (78.28%). At measured GFR < 45 ml/min/1.73 m, the BPNN has highest accuracy P30 (70.69%), and highest precision IQR (12.46 ml/min/1.73 m). The biases of BPNN and BIS1 equations were similar (0.74 [-1.55-2.78] and 0.24 [-2.58-1.61], respectively), smaller than any other equation.

Conclusions: The novel BPNN tool is more accurate than the currently available creatinine-based GFR estimation equations in an older population and could be recommended for routine clinical use.