Machine Learning-Based Prediction of Treatment Success, Postoperative Pain, and Analgesic Use after Full Pulpotomy in Permanent Molars with Symptomatic Irreversible Pulpitis.

INTRODUCTION

This study aimed to develop and evaluate machine learning (ML) models for predicting treatment success, postoperative pain, and analgesic use following full pulpotomy in permanent molars with symptomatic irreversible pulpitis.

METHODS

This retrospective multicohort study included 214 molars with a minimum 2-year follow-up. ML models, including logistic regression, support vector machines, and random forest, were developed for classification tasks, whereas ridge regression was used for postoperative pain prediction. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, F1 score, and calibration analysis with Brier scores. Internal validation was performed using a training-test split and cross-validation.

RESULTS

Predictive performance was highest for treatment success, for which classification models showed modest discriminative ability (AUC, 0.68-0.75), with logistic regression achieving the best performance (AUC = 0.75). Bleeding time emerged as the most influential predictor of treatment success. Prediction of analgesic use was limited (F1 score = 0.19), likely owing to class imbalance, and prediction of postoperative pain was weak (R2 = 0.21), indicating low explanatory capacity. Calibration analysis using the Brier score showed moderate agreement between predicted and observed outcomes.

CONCLUSIONS

When based on routinely available clinical variables, ML models showed the greatest predictive value for treatment success, whereas prediction of analgesic use and postoperative pain remained limited. Bleeding time may represent a clinically accessible variable with potential prognostic relevance. Although these models are not suitable for standalone clinical decision-making, they may support risk stratification, particularly for treatment success.