Abstract
AIM: This prognostic study aims to develop a machine learning (ML) survival model for estimating the longevity (success and survival rate) of restorations in endodontically treated teeth (ETT). METHODOLOGY: Data were consolidated from four controlled clinical trials conducted in the Netherlands and Brazil, involving 424 patients and 618 restorations with up to 17 years of follow-up. The evaluated predictive models included Gradient Boosting Survival, Random Survival Forests and Survival Support Vector Machine. The dataset was split into 70% for training and 30% for testing. Hyperparameter tuning was optimised via 10-fold cross-validation with 50 iterations using hyperopt. Performance was assessed through the time-dependent area under the ROC curve (AUC), concordance index (C-index), inverse probability of censoring weights (IPCW C-index) and time-dependent Brier score. RESULTS: The Gradient Boosting Survival model achieved the highest AUC mean (0.83, 95% confidence interval [CI], 0.81-0.78), C-index (0.80), IPCW C-index (0.78) and Brier score (0.06) for survival rate predictions, maintaining predictive stability over time. For success rate, the Random Survival Forest model outperformed others (AUC = 0.73, 95% CI [0.70-0.75]), C-index (0.66), IPCW C-index (0.64) and Brier score (0.14). SHAP analysis identified patient age and tooth type as having the highest variable importance for survival, while the dentist's experience was critical for success outcomes. Fairness analysis revealed performance disparities across sexes and countries in the models. CONCLUSIONS: The models demonstrated high predictive performance, mainly in survival rate prediction. ML models show promise for developing a robust, data-driven framework to evaluate success and survival outcomes in ETT.