Estimation of renal scarring in children with lower urinary tract dysfunction by utilizing resampling technique and machine learning algorithms

Aim: Classical database methods may be inadequate for large data sets accumulating continuously. Machine learning (ML), one of the main subsets of artificial intelligence, may solve this problem and find the best solution for future problems by gaining experience from the present data in medical studies. A method that may show the correlation between clinical findings and renal scarring (RS) with high accuracy in patients with lower urinary tract dysfunction (LUTD) is needed. In this study, the aim is to establish a model for the prediction of RS in children with LUTD by using ML. Methods: Patients older than three years of age (n=114) who needed urodynamic study were included in the study. There were 47 variables in the data set. Variables such as symptomatic urinary tract infection, vesicoureteral reflux, bladder trabeculation, bladder wall thickness, abnormal DMSA scintigraphy, and the use of clean intermittent catheterization were recorded. Several ML techniques (MLT) were applied to estimate RS. Results: As a result of the comparisons, the highest accuracy rate according to the confusion matrix was obtained by the Extreme Gradient Boosting (XGB) algorithm (91.30%). In the balanced (SMOTE) data set, the highest accuracy rate was obtained by the Artificial Neural Network (ANN) algorithm (90.63%). According to the Receiver Operating Characteristic (ROC), the highest success rate was obtained by the ANN algorithm in the balanced (SMOTE) data set (90.78%). Conclusion: High accuracy rates obtained by MLT may suggest that MLT might provide a faster and accurate evaluation process in the estimation of RS in patients with LUTD.