Computational Model for Predicting the Chance of Early Resolution in Children With Vesicoureteral Reflux

Purpose: Minimally invasive treatment options and concern regarding long-term antibiotics have increased emphasis on predicting the chance of early vesicoureteral reflux resolution. Computational models, such as artificial neural networks, have been used to assist decision making in the clinical setting using complex numeric constructs to solve multivariable problems. We investigated various computational models to enhance the prediction of vesicoureteral reflux resolution. Materials and Methods: We reviewed the records of 205 children with vesicoureteral reflux, including 163 females and 42 males. In addition to reflux grade, several clinical variables were recorded from the diagnostic visit. Outcome was noted as resolved or unresolved at 1 and 2 years after diagnosis. Two separate data sets were prepared for the 1 and 2-year outcomes, sharing the same input features. The data sets were randomized into a modeling set of 155 and a cross-validation set of 50. The model was constructed with several constructs using neUROn++, a set of C++ programs that we developed, to best fit the data. Results: A linear support vector machine was found to have the highest accuracy with a test set ROC curve area of 0.819 and 0.86 for the 1 and 2-year models, respectively. The model was deployed in JavaScript® for ready availability on the Internet, allowing all input variables to be entered and calculating the odds of 1 and 2-year resolution. Conclusions: This computational model allowed the use of multiple variables to improve the individualized prediction of early reflux resolution. This is a potentially useful clinical tool regarding treatment decisions for vesicoureteral reflux.