Predicting orchiectomy in testicular torsion using hybrid machine learning and explainable AI: a web-based clinical...

To develop a high-accuracy prediction model using hybrid machine learning (ML) and explainable artificial intelligence (XAI) techniques for distinguishing between orchiectomy and detorsion in testicular torsion (TT), and to create an interactive web application and nomogram for clinical use.

Data from 117 patients who underwent surgical treatment for TT at our clinic between January 2005 and June 2025 were retrospectively analyzed (detorsion: 83, orchiectomy: 34). From twenty initial features, seven optimal features were selected using a hybrid Particle Swarm Optimization-Grey Wolf Optimization (PSO-GWO) algorithm. Class imbalance was addressed using SVMSMOTE. Four ensemble learning algorithms were compared, and CatBoost was selected as the final model. SHAP, LIME, Partial Dependence Plot (PDP), and Individual Conditional Expectation (ICE) analyses were performed for model interpretability. A seven-feature nomogram was developed using logistic regression. A two-module web application was created using Python Dash.

PSO-GWO identified seven optimal features (Age, Symptom Duration, PDW, PCT, MPV, MLR, SII). CatBoost demonstrated the highest performance (AUC: 0.923, Accuracy: 89.5%). XAI analyses identified symptom duration as the strongest predictor (SHAP: 52.72%), followed by MLR (14.24%) and PDW (12.09%). PDW emerged as a novel biomarker with minimal prior investigation in the literature yet demonstrated strong discriminatory power (inverse relationship, cut-off < 17.9 fL). PDP analysis revealed a distinct inflection point at approximately 7 h, corroborating the ROC cut-off values. ICE analysis revealed substantial individual heterogeneity (score > 0.36). The nomogram demonstrated good discrimination (AUC: 0.818) and excellent calibration (Hosmer-Lemeshow p = 0.556). A simplified three-feature risk scoring system (0-3 points) was developed. The web application (medicalinformaticsttrc.adiyaman.edu.tr) provided real-time risk calculation (< 50 ms).

The CatBoost model developed with hybrid PSO-GWO algorithm and XAI techniques achieved high accuracy in predicting orchiectomy in TT. PDW was identified as a novel, practical, and cost-effective biomarker. The developed web application and risk scoring system may support preoperative counseling, surgical preparedness, and clinical decision-making in the emergency department. The "detorsion" outcome indicates intraoperative non-removal of the testis and does not confirm long-term testicular viability. This exploratory model requires external validation and multicenter prospective studies before clinical adoption.