Predicting longevity of dental restorations using artificial intelligence: a survival machine learning approach

Abstract

Objective: to develop and evaluate machine learning survival models for predicting restoration longevity. Methods: This retrospective prognostic prediction study used electronic patient records from general dental practices in the Netherlands. The outcome was restoration failure, defined as any subsequent restorative intervention affecting the same restored surfaces. Candidate predictors included patient-, tooth-, and restoration-level. Missing data were handled using Multiple Imputation by Chained Equations. Categorical variables were one-hot encoded, and continuous variables were normalized using z-score. Data were split at the individual level into training (80%) and testing (20%) sets. Three survival machine learning algorithms were evaluated: Gradient Boosting Survival (GBS), Random Survival Forests (RSF), and Survival Support Vector Machine (SSVM). Performance was assessed using time-dependent AUC, concordance index (C-index), IPCW C-index, and Brier score. Results: A total of 115,522 restorations from 23,561 individuals were included. In the test set, no significant differences between the models were observed. SVM achieved an AUC=0.75 (95%CI:0.72-0.77), C-index=0.74, and          IPCW C-Index=0.74. RSF achieved AUC=0.75 (95%CI:0.72-0.77), C-index=0.74, IPCW C-Index=0.73, and a Brier Score of 0.02. GBS achieved an AUC=0.71 (95%CI:0.68-0.74), C-index=0.71, IPCW C-Index=0.69, and a Brier Score of 0.02. Fairness analysis showed similar predictive performance across sex and socioeconomic groups, with low parity differences (0.02–0.05), disparate impact values close to 1 (0.93–1.09), and low equalized odds differences (0.04–0.10), suggesting limited algorithmic bias. SHAP analysis identified molar, general health, the patient age, and prosthesis use as the most influential predictors. Conclusions: Machine learning survival models demonstrated good predictive performance and equitable behavior