IP58-13 DEVELOPMENT AND VALIDATION OF PROCEDURE-SPECIFIC MACHINE LEARNING MODELS FOR PREDICTING PROSTHESIS INFECTION RISK: A CLINICAL DECISION SUPPORT TOOL

Alexandria M. Hertz; Jonathan Seaman; Maali La France; Aziz Shabaan; Ryan Hanson; Maia Van Dyke; Steven Hudak

doi:10.1097/01.JU.0001191628.19015.5e.13

INTRODUCTION AND OBJECTIVES: Prosthesis-associated infections represent a significant complication following urologic implant procedures. Current risk stratification relies on subjective assessment without standardized, procedure-specific prediction tools. We wished to develop an internally validated procedure-specific machine learning model to predict infection risk and create a clinical decision support framework. METHODS: We analyzed 1,646 consecutive urologic prosthetic procedures performed between 2013-2025: 897 penile prosthesis (PP) implantations (including malleable devices) and 742 artificial urinary sphincter (AUS) placements. We captured 23 clinical variables including patient demographics, comorbidities, operative factors, and antibiotic prophylaxis patterns. The primary outcome was device-associated infection requiring explant post-operatively. We developed separate machine learning models for each procedure type. Due to low infection rates, we implemented Random Over-Sampling Examples (ROSE) and tested multiple algorithms including Random Forest, XGBoost, and logistic regression. Model calibration was achieved using Platt scaling and isotonic regression. Performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC), sensitivity, specificity, and calibration metrics with 5-fold cross-validation and temporal data splitting (80% training, 20% testing). RESULTS: Overall infection rate was 2.0% (33/1,646), different rates for PP 2.34% (21/897) versus AUS 1.62% (12/742). Procedure-specific modeling demonstrated AUC for PP of 0.628 and 0.598 for AUS. Risk factor importance varied significantly by procedure type. For PP, coronary artery disease (Risk Ratio [RR] 1.59, p<0.05), elevated BMI (mean difference +1.0), and longer operative time (+3.2 minutes) were primary predictors. For AUS, radiation history emerged as the strongest predictor (RR 1.95, p<0.01), followed by diabetes (RR 1.87, p<0.05). Model calibration successfully reduced prediction-reality gaps from >20% to <2%, enabling clinically meaningful risk stratification. We established four risk categories (Low <1.5%, Moderate 1.5-4.5%, High 4.5-10%, Very High >10%) with corresponding evidence-based clinical protocols and developed a web-based clinical decision support interface. CONCLUSIONS: This represents the first internally validated, procedure-stratified prediction tool for prosthetic infection risk. External validation studies at multiple institutions and prospective trials are warranted to establish clinical utility and impact on patient outcomes.

IP58-13 DEVELOPMENT AND VALIDATION OF PROCEDURE-SPECIFIC MACHINE LEARNING MODELS FOR PREDICTING PROSTHESIS INFECTION RISK: A CLINICAL DECISION SUPPORT TOOL

View Online

Abstract

Details

Metrics