Prediction of severe erectile dysfunction after penile fracture repair: machine learning analysis results from the reconstruction and trauma working group of the society of urological surgery (RAT-SUS).

Background: Erectile dysfunction (ED) is a significant complication following penile fracture repair, and early prediction is critical for clinical management. Aim: To evaluate the effectiveness of machine learning (ML) algorithms in predicting the development of severe ED after penile fracture repair and to identify complex risk factors beyond the scope of traditional statistical methods. Methods: A retrospective analysis was conducted using data from 547 patients who underwent surgical repair for penile fracture between January 2020 and June 2024 at 23 urology centers affiliated with the Reconstructive Urology and Trauma Study Group of the Urological Surgery Society. Patients were categorized into two groups based on their International Index of Erectile Function-5 scores at six months postoperatively: severe ED (+) (≤7) and ED (-) (>7). Eleven different ML classifiers were evaluated to determine the most predictive models. Four distinct resampling techniques were employed to address class imbalance in the dataset. Feature importance analysis was also performed to identify the most influential variables contributing to ED risk. Outcomes: This study was conducted to enable the early identification of patients at high risk of developing severe ED following penile fracture surgery. Results: Logistic Regression, Gaussian Naive Bayes, and Linear Support Vector Machine emerged as the best-performing algorithms on the original dataset, with Area Under the Curve (AUC) scores of 0.81, 0.78, and 0.76, respectively. On the Synthetic Minority Over-sampling Technique (SMOTE)-resampled dataset, Quadratic Discriminant Analysis (QDA) achieved an AUC of 0.85, while the Artificial Neural Network (ANN) reached an AUC of 0.84. On the SMOTE-resampled dataset, QDA achieved a ROC-AUC of 0.85 (95% CI: 0.75-0.93), whereas on the SMOTE-Tomek Link-resampled dataset, the ANN attained a ROC-AUC of 0.84 (95% CI: 0.71-0.94). The most critical predictors of severe ED were age, comorbidities, tunical tear length, and time to surgery. Urethral injuries were not significant contributors, as all were minor and managed conservatively without urethroplasty. Clinical Implications: Integration of ML-based prediction models into clinical workflows could support early risk stratification and individualized patient care, ultimately improving postoperative functional outcomes. Strengths and Limitations: This study benefits from a large, multicenter dataset and a comparative analysis of multiple ML algorithms. However, its retrospective nature and inter-center variability in data reporting may limit generalizability. Conclusion: ML algorithms are effective and reliable tools for predicting severe ED after penile fracture repair and may enhance personalized postoperative management. Eliminating class imbalance in the data with resampling techniques improves model performance.