Explainable machine learning for perioperative surgical site infection risk enrichment after operative treatment of...

Surgical site infection (SSI) remains a major complication after operative treatment of closed pilon fractures. Prediction tools may help identify patients who warrant closer perioperative wound surveillance, but externally validated and clinically interpretable models remain limited. We aimed to develop and externally validate explainable machine-learning models for perioperative SSI risk enrichment after operative treatment of closed pilon fractures.

This retrospective multicenter study included 1,876 patients treated at the internal center between January 2020 and December 2024 and 359 patients from an independent external center treated between August 2024 and September 2025. The internal dataset was stratified into a training set and an internal test set at a 7:3 ratio. Candidate predictors included demographic, injury-related, perioperative, laboratory, and inflammatory-metabolic variables. LASSO regression was used for feature selection, followed by logistic regression, decision tree, random forest (RF), XGBoost, and naïve Bayes models. Thresholds were optimized using out-of-fold cross-validation predictions. Model performance was evaluated by ROC-AUC, PR-AUC, sensitivity, specificity, precision, F1 score, balanced accuracy, calibration indices, Brier score, bootstrap 95% confidence intervals, and decision-curve analysis. A preoperative-only RF sensitivity model was also constructed.

The internal cohort included 74 SSI events (3.9%), and the external cohort included 11 SSI events (3.1%). LASSO retained 16 predictors. The RF model showed the highest overall discrimination, with ROC-AUCs of 0.899 (95% CI, 0.831-0.953) in the internal test set and 0.902 (95% CI, 0.707-0.991) in the external cohort. Sensitivity was 0.294 (95% CI, 0.077-0.533) internally and 0.636 (95% CI, 0.333-0.909) externally. RF specificity was high, at 0.987 (95% CI, 0.976-0.996) and 0.974 (95% CI, 0.957-0.989), respectively. RF Brier scores were 0.026 and 0.022, calibration slopes were 1.915 and 1.668, and observed-to-expected ratios were 0.736 and 0.762 in the internal and external cohorts. The preoperative-only RF model achieved ROC-AUCs of 0.884 and 0.905 in the internal and external cohorts, respectively. Decision-curve analysis showed positive RF net benefit across threshold probabilities of 0.01-0.38 internally and 0.01-0.49 externally.

The present machine-learning framework showed favorable discrimination for perioperative SSI risk enrichment after closed pilon fracture surgery, with RF serving as the main high-specificity model. However, limited RF sensitivity and calibration uncertainty suggest that it should be viewed as a preliminary targeted-surveillance tool rather than a rule-out or implementation-ready system. Further recalibration and prospective multicenter validation are needed before routine clinical use.