Machine learning models for predicting postoperative complications following mandibular third molar surgery: Development, validation, and explainable AI insights.

The surgical removal of impacted mandibular third molars is a routine procedure in oral and maxillofacial surgery but is consistently associated with significant postoperative morbidity. Traditional indices of surgical difficulty rely mainly on positional and radiographic features, overlooking intraoperative and operator-related influences. This retrospective cohort study analyzed 472 cases of mandibular third molar extractions to develop and validate machine learning models for predicting postoperative complications. Demographic, radiographic, and intraoperative variables were included, and outcomes assessed were pain, swelling, trismus, alveolar osteitis, infection, paraesthesia, bleeding, and delayed wound healing. Five algorithms were compared, with Random Forest and XGBoost demonstrating superior predictive performance (AUC 0.91 and 0.89, respectively) compared with logistic regression (AUC 0.74). SHapley Additive exPlanations (SHAP) identified surgical duration, impaction type, root morphology, operator experience, and bone guttering as the most influential predictors. These findings highlight the potential of explainable artificial intelligence to provide clinically interpretable and individualized risk stratification, supporting better triage, patient counseling, and training allocation. While single-center and retrospective design limit generalizability, this study demonstrates that machine learning can enhance prediction of third molar surgery outcomes beyond conventional methods and provides a foundation for prospective multicenter validation.