Back to 2026 Abstracts
Predicting Post-Operative IPP Infections: Machine Learning Risk Stratification and Validation
Muhammad Hammad, MBBCh1, Martin Gross, MD2, Nicholas Moll, MD2, Samuel Ivan, MD3, Elia Chawareb, MD1, Yash Kadakia, MD1, Faysal Yafi, MD1, Thairo Pereira, MD4, Jacob Good, MD5, Helen Bernie, MD5, Nicklas Sarantos, MD6, Aaron Lentz, MD6, Alfredo Suarez-Sarmiento, Jr., MD7, Paul Perito, MD7, Joshua Schammel, MD8.
1UC Irvine, Irvine, CA, USA, 2Dartmouth Hitchcock Medical Center, Lebanon, NH, USA, 3Fox Chase, Philadelphia, PA, USA, 4University of Rochester, Rochester, NY, USA, 5Indiana University, Indianapolis, IN, USA, 6Duke University, Durham, NC, USA, 7Perito Urology, Coral Gables, FL, USA, 8Albany Medical College, Albany, NY, USA.
INTRODUCTION:Post-operative infection following inflatable penile prosthesis (IPP) surgery remains one of the most serious complications despite advances in antiseptic techniques. Current infection risk-assessment relies on surgeon experience rather than objective tools. We developed and validated a machine learning-based model using a unified complication taxonomy to enable preoperative infection risk stratification.
METHODS:A total of 5,815 multicenter IPP cases (primary, revision, and infection-salvage) were analyzed. Unstructured operative entries were standardized into a 38-feature taxonomy encompassing demographics, comorbidities, surgical variables, prosthesis type, and intraoperative details. Among these, 117 cases (2%) were complication-positive for post-operative infection. We compared 80 model configurations (10 classifiers × 4 resampling strategies × 2 data splits). Model performance was evaluated using macro-averaged precision, recall, F1 score, and precision-recall area under the curve (PR-AUC). Feature interpretability was assessed with SHAP (Shapley Additive Explanations).
RESULTS:Gradient Boosting with Random Undersampling (RUS) achieved the best standard-split performance (Precision = 0.52, Recall = 0.75, F1 = 0.46, PR-AUC = 0.10). Bernoulli Naïve Bayes without resampling performed best on temporal validation (Precision = 0.52, Recall = 0.69, F1 = 0.48, PR-AUC = 0.06). SHAP analysis (Fig 1) identified diabetes, coronary/peripheral vascular disease, and history of prior IPP infection as the strongest positive predictors of infection, with maximum SHAP contributions of +0.21, +0.67, and +0.83, respectively. Additional risk factors included intra-operative complications and older age. Protective associations were observed for antibiotic impregnated LGX use and combined peri-operative antifungal + vancomycin prophylaxis.
CONCLUSIONS:A gradient-boosting model integrating pre- and intra-operative variables achieved 75% sensitivity for post-operative infection prediction, supporting its role as a preoperative screening tool. SHAP interpretability highlights patient comorbidities and infection history as dominant drivers of infection risk, while combination antifungal-antibiotic prophylaxis may confer
measurable protection. External validation is underway to optimize model calibration and clinical deployment.
Back to 2026 Abstracts
