Back to 2026 Display Posters
When Prediction Changes Prescription: Machine-Learning Estimates and Physician Decisions on Antibiotic Prophylaxis for Vesicoureteral Reflux
HsinHsiao Scott Wang, MD, MPH, MBAn1, Michael Li, MBAn, PhD2, Agni Orfanoudaki, PhD3, Carlos Estrada, MD, MBA1.
1Boston Children's Hospital, Boston, MA, USA, 2Harvard Business School, Boston, MA, USA, 3Oxford Saïd Business School, Oxford, United Kingdom.
Background: Antibiotic prophylaxis (AP) for children with vesicoureteral reflux (VUR) remains debated, in part because randomized trial results have been interpreted differently across clinical practice. While machine-learning (ML) models may support individualized risk prediction and treatment-benefit estimation, it remains unclear how physicians incorporate such information into AP decision-making. We aimed to evaluate whether ML-estimated AP benefit influences physician recommendations for AP in children with VUR.
Methods: We conducted a physician survey among pediatric urologists using five clinical scenarios involving children aged 2 months to 6 years with confirmed VUR. For each scenario, respondents were shown ML-estimated AP benefit and asked whether they would recommend AP. To establish baseline prescribing patterns for each physician, we retrospectively reviewed new VUR diagnoses from 2023-2025, excluding patients with complex anatomy. Baseline data included demographics, initial urinary tract infection characteristics, VUR grade, bowel/bladder dysfunction or constipation history, and AP assignment. Optimal multivariable matching was used to account for differences between patients in baseline practice and survey scenarios.
Results: Fifteen physicians were included. At baseline, physicians recommended AP for 89% of children with VUR (95% CI, 77%-98%). After presentation of ML-estimated AP benefit, physicians recommended AP less frequently, with an average absolute reduction of 22% compared with baseline practice (95% CI, 3%-41%; p=0.024). At the physician level, male physicians demonstrated a greater shift away from AP recommendation compared with female physicians (93% vs 69%; p=0.017). At the patient level, reduction in AP recommendation was most pronounced among scenarios involving low-grade VUR (100% vs 57%; p=0.027) and no history of constipation or bowel/bladder dysfunction (92% vs 61%; p=0.003). Physicians reporting high familiarity with ML showed a trend toward greater reduction in AP use after viewing ML-estimated benefit, although this did not reach statistical significance (100% vs 55%; p=0.056).
Conclusion: Although urologists strongly favored AP for children with VUR at baseline, presentation of ML-estimated AP benefit was associated with a significant shift toward less AP use and more individualized decision-making. The greatest changes occurred in clinical scenarios traditionally associated with lower risk, including low-grade VUR and absence of bowel/bladder dysfunction or constipation. Although the ML output in this study was theoretical, these findings suggest that patient-specific estimates of treatment benefit may meaningfully influence physician decisions and should inform the design, presentation, and clinical integration of ML-based decision-support tools in pediatric VUR care.
Back to 2026 Display Posters
