MP64-03 SELECTING CHILDREN WITH VUR WHO ARE MOST LIKELY TO BENEFIT FROM ANTIBIOTIC PROPHYLAXIS: APPLICATION OF MACHINE LEARNING TO RIVUR DATA

Abstract

INTRODUCTION AND OBJECTIVES: RIVUR trial found that antibiotic prophylaxis (AP) reduced the risk of recurrent UTI (rUTI) by 50% in VUR. However, there may be subgroups in whom AP is either more or less effective. If they were identifiable, AP could be used more selectively. We sought to develop a machine learning model to identify such subgroups in RIVUR. METHOD(S): RIVUR data were randomly split into train/test sets in 4:1. Two optimal classification tree models were built for AP and non-AP subgroups to predict rUTI risk in both scenarios. The test set was then used to validate actual rUTI events and the effectiveness of AP. Two predicted probabilities were generated from each prediction model and AP was assigned according to cutoffs for AP treatment effectiveness defined by rUTI risk difference. RESULT(S): 607 patients (558 female/49 male, median age 12 mo) were included. Predictor variables in the model included VUR grade, serum creatinine, race/sex, prior UTI symptoms (fever, dysuria), and weight percentiles. The AUC of the joint prediction model of rUTI by treatment arms (AP vs placebo, Figure 1) was 0.81(0.75-0.86). Figure 2 compares the rUTI rate for randomly assigned AP (dashed lines) versus assignment by the model using risk reduction cutoffs (solid line). Using the model to assign AP based on >10% rUTI risk reduction, similar population efficacy can be achieved by treating only 40% of VUR patients compared to treating everyone. In a 120-patient test set, 51 patients had actual AP randomization consistent with our model recommendation (AP if absolute rUTI risk reduction > 10%). Actual rUTI incidence was significantly lower among those whose AP randomization was consistent with our model suggestion, compared to those whose AP assignment differed from model suggestion (2 vs 12.9%, p=0.03). CONCLUSION(S): Our predictive model using machine learning algorithms identifies VUR patients who are more likely to benefit from AP, which would allow more selective, personalized use of AP with maximal benefit, while minimizing use in those least likely to benefit.