Machine Learning Models for Predicting Periapical Healing after Primary Nonsurgical Endodontic Treatment

This study aimed to train and internally evaluate machine learning models to classify periapical healing as favorable (FAV) or unfavorable (UNF) after primary nonsurgical endodontic treatment, emphasizing failure-oriented evaluation.

This retrospective study included 276 single-tooth treatments performed by endodontists, with clinical-radiographic outcomes adjudicated after 12-18 months. Logistic regression, random forest (RF), and radial-basis-function support vector machine (SVM-RBF) models were developed using Scikit-learn pipelines. Validation used a stratified 80/20 train-test split and stratified 5-fold cross-validation. Class imbalance was addressed by random oversampling and synthetic minority oversampling applied only to training data. UNF healing was the positive class. Performance was assessed using ROC-AUC and precision-recall area under the curve for UNF healing (PR-AUC_UNF); a threshold-adjusted UNF-alert strategy was explored.

UNF healing occurred in 18.8% of cases. A majority-class baseline achieved 80.4% accuracy but 0% UNF sensitivity. At default thresholds, RF showed the highest discrimination (PR-AUC_UNF = 0.359; ROC-AUC = 0.724) but only 27.3% sensitivity. With imbalance handling, SVM-RBF achieved PR-AUC_UNF up to 0.535 and sensitivity up to 54.5%. Threshold adjustment increased sensitivity to 90.9% while reducing specificity to 48.9% and positive predictive value to 30.3%. Tooth type and baseline lesion size had the highest aggregated RF feature-importance scores.

Accuracy alone may obscure clinically relevant failures in imbalanced datasets. Threshold adjustment increased UNF detection at the cost of false-positive alerts. These internally evaluated findings are preliminary and require external validation before clinical use.