Prediction of Surgical Intervention in Acute Knee Trauma: A Focus on Threshold-Specific Performance and Clinical Decision Utility

Background: Acute knee trauma is a common reason for emergency department visits, yet early identification of patients requiring surgical intervention remains challenging. Most existing prediction studies focus on discrimination metrics and provide limited guidance for clinical decision-making. Methods: We conducted a retrospective study of 905 patients presenting to the emergency department with acute knee trauma. Prediction models were developed using logistic regression, random forest, and extreme gradient boosting (XGBoost) based on routinely available clinical variables. Model performance was evaluated in terms of discrimination (AUROC, AUPRC), calibration, and clinical utility. Threshold-specific performance metrics and decision curve analysis were used to assess clinical applicability, and patients were stratified into risk groups based on predicted probabilities. Results: Among 905 patients, 163 (18.0%) underwent surgical intervention. Logistic regression and random forest demonstrated comparable performance (AUROC 0.748 and 0.744, respectively), whereas XGBoost showed lower discrimination (AUROC 0.632). Calibration was acceptable overall but less stable at higher predicted probabilities. Threshold-specific analysis demonstrated meaningful trade-offs between sensitivity and specificity across probability thresholds. Decision curve analysis showed that the model provided greater net benefit than default strategies within a threshold range of approximately 0.05–0.25. Risk stratification showed increasing surgical rates across risk groups, although the degree of separation was modest. Conclusions: Prediction models based on routinely available clinical variables can support early risk assessment in acute knee trauma. Their clinical usefulness depends on threshold-specific evaluation and decision-analytic approaches rather than overall performance metrics alone. These findings highlight the importance of interpreting prediction models within a clinical decision-making framework to facilitate real-world application.