The use of artificial neural networks in decision support in vesicoureteral reflux treatment

To develop a prediction model based on artificial neural networks (ANN) for the treatment selection in vesicoureteral reflux (VUR).

A total of 96 children with VUR (145 ureteric units (UU)) were treated at our institution during 2004-2006. An ANN based on quick propagation architecture was created with the commercially available software package. The patients' age and sex, the cause and grade of VUR, the affected ureter, the type of treatment (conservative, subureteric injection, or open surgery), existence of renal scar on DMSA, follow-up times and the number of injections were used as variables. These data were also transferred to a statistical software package and regression analysis was done.

In all, 105 UU showed no reflux, 5 UU showed improvements in reflux grade (considered only in the conservative management group), and the remaining 35 UU showed persistence. In the training group (n = 99), ANN showed 98.5% sensitivity, 92.5% specificity, 97% positive predictive value, and 96% negative predictive value in predicting treatment outcome.

We have demonstrated that ANN can accurately predict the resolution of VUR, and thus could be useful in daily clinical practice. This approach would allow urologists to aid in the decision-making process of VUR treatment.