A comparison of the performances of an artificial neural network and a regression model for GFR estimation.

BACKGROUND: Accurate estimation of glomerular filtration rate (GFR) is important in clinical practice. Current models derived from regression are limited by the imprecision of GFR estimates. We hypothesized that an artificial neural network (ANN) might improve the precision of GFR estimates. STUDY DESIGN: A study of diagnostic test accuracy. SETTING & PARTICIPANTS: 1,230 patients with chronic kidney disease were enrolled, including the development cohort (n=581), internal validation cohort (n=278), and external validation cohort (n=371). INDEX TESTS: Estimated GFR (eGFR) using a new ANN model and a new regression model using age, sex, and standardized serum creatinine level derived in the development and internal validation cohort, and the CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) 2009 creatinine equation. REFERENCE TEST: Measured GFR (mGFR). OTHER MEASUREMENTS: GFR was measured using a diethylenetriaminepentaacetic acid renal dynamic imaging method. Serum creatinine was measured with an enzymatic method traceable to isotope-dilution mass spectrometry. RESULTS: In the external validation cohort, mean mGFR was 49±27 (SD) mL/min/1.73 m2 and biases (median difference between mGFR and eGFR) for the CKD-EPI, new regression, and new ANN models were 0.4, 1.5, and -0.5 mL/min/1.73 m2, respectively (P<0.001 and P=0.02 compared to CKD-EPI and P<0.001 comparing the new regression and ANN models). Precisions (IQRs for the difference) were 22.6, 14.9, and 15.6 mL/min/1.73 m2, respectively (P<0.001 for both compared to CKD-EPI and P<0.001 comparing the new ANN and new regression models). Accuracies (proportions of eGFRs not deviating >30% from mGFR) were 50.9%, 77.4%, and 78.7%, respectively (P<0.001 for both compared to CKD-EPI and P=0.5 comparing the new ANN and new regression models). LIMITATIONS: Different methods for measuring GFR were a source of systematic bias in comparisons of new models to CKD-EPI, and both the derivation and validation cohorts consisted of a group of patients who were referred to the same institution. CONCLUSIONS: An ANN model using 3 variables did not perform better than a new regression model. Whether ANN can improve GFR estimation using more variables requires further investigation.

Estimation of glomerular filtration rate by a radial basis function neural network in patients with type-2 diabetes mellitus.

BACKGROUND: Accurate and precise estimates of glomerular filtration rate (GFR) are essential for clinical assessments, and many methods of estimation are available. We developed a radial basis function (RBF) network and assessed the performance of this method in the estimation of the GFRs of 207 patients with type-2 diabetes and CKD. METHODS: Standard GFR (sGFR) was determined by (99m)Tc-DTPA renal dynamic imaging and GFR was also estimated by the 6-variable MDRD equation and the 4-variable MDRD equation. RESULTS: Bland-Altman analysis indicated that estimates from the RBF network were more precise than those from the other two methods for some groups of patients. However, the median difference of RBF network estimates from sGFR was greater than those from the other two estimates, indicating greater bias. For patients with stage I/II CKD, the median absolute difference of the RBF network estimate from sGFR was significantly lower, and the P50 of the RBF network estimate (n = 56, 87.5%) was significantly higher than that of the MDRD-4 estimate (n = 49, 76.6%) (p < 0.0167), indicating that the RBF network estimate provided greater accuracy for these patients. CONCLUSIONS: In patients with type-2 diabetes mellitus, estimation of GFR by our RBF network provided better precision and accuracy for some groups of patients than the estimation by the traditional MDRD equations. However, the RBF network estimates of GFR tended to have greater bias and higher than those indicated by sGFR determined by (99m)Tc-DTPA renal dynamic imaging.