Using Interpretable Machine Learning Model to Predict Orchiectomy after Testicular Torsion

Background This study aimed to develop and evaluate machine learning (ML) models for the preoperative prediction of orchiectomy in patients with testicular torsion. Methods We conducted a retrospective analysis of 204 cases of suspected testicular torsion managed with surgical exploration between January 2003 and April 2025. The patient cohort was partitioned via stratified sampling into a training dataset (70%) and a holdout testing dataset (30%). Initially, Least absolute shrinkage and selection operator (LASSO) regression was employed to identify six optimal predictors. Subsequently, five ML models were developed on these predictors and their performance was validated on the testing set. Model efficacy was evaluated on the testing set using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Finally, we utilized Shapley Additive Explanations (SHAP) to assess model interpretability and quantify the impact of each feature. Results Among the five ML models, LightGBM model exhibited the superior predictive outcomes on the testing set. Its performance was quantified by an AUC of 0.879 (95% CI: 0.775–0.983) and an accuracy of 0.806. An examination of the model's inner workings using SHAP values determined the relative importance of each feature. The results revealed that symptom duration, degree of torsion, abdominal pain, fibrinogen, monocyte, and lymphocyte-to-monocyte ratio (LMR) were the primary drivers of its predictive power. Conclusion To predict orchiectomy in testicular torsion before surgery, we designed and validated several machine learning models. Among them, the LightGBM model using six clinical predictors demonstrated superior performance.