Predicting the Severity of Postoperative Symptoms Following Mandibular Third Molar Extractions Using Machine Learning...

This study aims to develop and externally validate machine-learning models that effectively predict the risk and severity of postoperative symptoms one week following mandibular third molar extractions.

This retrospective cohort study included 321 patients (18-35 years old) who underwent lower third-molar surgery. Demographics, Pell-Gregory vertical (PGV) and Pell-Gregory level (PGL) classifications, surgical variables, and day-7 pain visual analogue scale (VAS) were recorded for all participants. The data were randomly divided into training (70%) and validation (30%) datasets. Five machine-learning algorithms-Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Decision Tree (DT), and Neural Network (NNET)-were developed using nested cross-validation. Model performance was assessed through area under the receiver operating characteristic (AUROC) values, Brier scores, and calibration slopes, with a nomogram constructed from the best-performing model.

GBM achieved the highest discrimination on the validation dataset with an AUROC of 0.687 (95% CI, 0.624-0.744), followed by the Neural Network (AUROC = 0.677). The GBM model yielded a calibration slope of 0.98 and a Brier score of 0.225, indicating excellent predictive accuracy. However, the top six predictors were found to be operative time, mouth opening, PGV, PGL, smoking, and preoperative symptoms. The GBM model, which underlies the nomogram, achieved an area under the curve (AUC) value of 0.666, indicating its discrimination capability. Additionally, the calibration curve confirmed the model's accuracy, and the decision curve analysis (DCA) suggested that the nomogram provides clinically promising potential for effective risk stratification.

A GBM-based nomogram provides moderate yet clinically useful discrimination for healthy adults aged 18-35 years at risk for severe early symptoms after third-molar extraction. However, this approach requires external validation in older or medically complex patients before it is recommended for clinical predictions.