Development and validation of a machine learning model to predict mandibular third molar impaction and associated...

Impacted mandibular third molars are common worldwide and may cause complications such as pericoronitis, cysts, and second molar caries. Conventional prediction methods rely on subjective radiographic interpretations and clinical judgment, which can vary among practitioners. Recent advances in machine learning (ML) provide opportunities to develop objective, data-driven models for clinical decision-making in third molar management.

To develop and validate a machine learning-based model for predicting mandibular third molar impaction and associated complications using demographic and radiographic variables.

This retrospective observational study evaluated 220 panoramic radiographs of patients aged 16-40 years. Collected variables included age, sex, tooth angulation, Pell and Gregory classification, depth of impaction, root development, ramus relationship, and proximity to the mandibular canal. Logistic regression, random forest, and XGBoost were trained on 70 % of the dataset and validated on 30 %. Model performance was assessed using AUC-ROC, accuracy, sensitivity, specificity, calibration, and Cohen's Kappa for agreement with expert judgment.

Impaction prevalence was 67.3 %. Significant predictors included mesioangular angulation, Pell and Gregory Class II, and incomplete root development (p < 0.001). XGBoost outperformed other models, achieving an AUC-ROC of 0.92, accuracy of 90.5 %, and Kappa of 0.82. Pericoronitis (26 %) and distal second molar caries (18 %) were the most common complications.

XGBoost demonstrated high predictive accuracy for mandibular third molar impaction and complications. As a probability-based decision-support tool, it can provide individualized risk estimates or binary classifications to assist clinicians in counseling, surveillance, and surgical decision-making.