ARTIFICIAL NEURAL NETWORKS FOR ESTIMATING GLOMERULAR FILTRATION RATE BY URINARY DIPSTICK FOR TYPE 2 DIABETIC PATIENTS

Abstract

Background: The result of a standard urinary dipstick from a patient with diabetes mellitus type 2 can be used to predict the estimated glomerular filtration rate (eGFR). We designed a multilayer perceptron (MLP) to investigate the possibility and optimal number of variables for the prediction. Methods: A total of 299 volunteers with diabetes mellitus type 2 were included. The blood and urine samples from volunteers were analyzed for blood sugar, glycated hemoglobin, serum creatinine, and urine chemistry. The urine chemistry was examined by a standard urinary dipstick. Volunteer age and gender and six test items of the dipstick were set as eight variables for this study. The eight variables were grouped and examined for the optimal combination. The eight variables from 232 of 299 volunteers were used to train an MLP for the optimal variables. The performance of trained MLP was validated by the data from 69 of 232 volunteers. Results: The optimal combination for variables was the six test items of the dipstick and volunteer age. The area under the curve (0.928), accuracy (0.879), sensitivity (0.83), and specificity (0.88) of the trained MLP were examined. Conclusions: The results demonstrate the eGFR prediction potential of the results of a urinary dipstick using this method.