Predicting restoration failures in primary and permanent teeth - A machine learning approach.

Abstract

Objective: Machine learning (ML) predictive models promise to handle complex data and deliver accurate predictions in the medical field. The aim of this study was to develop ML predictive models for posterior dental restorations failures in both primary and permanent teeth.

Methods: Data from two clinical datasets were used in this study, encompassing a Randomized Controlled Trial (RCT) for permanent teeth (CaCIA Trial) and a corresponding RCT for primary teeth (CARDEC 3). Models were developed using five different algorithms-Decision Tree, Random Forest, XGBoost, CatBoost and Neural Network-ensuring thorough cross-validation and calibration for predictive reliability. Clinical variables related to patients and teeth were considered as predictors. Model performances were assessed using accuracy, precision, recall, F1-score and ROC AUC, alongside SHAP plots for interpretability.

Results: In the primary teeth dataset, all models demonstrated acceptable performance with AUC values around 0.67-0.75 and a balanced trade-off between precision and recall. In contrast, the models applied to permanent teeth yielded less predictive ability, with AUC values ranging from 0.53 to 0.62.

Conclusion: Our results highlight how ML approaches effectively process intricate, multi-dimensional data related to restoration longevity, successfully integrating variables across patient characteristics, tooth properties, and diagnostic assessments within a unified analytical framework. Though promising as analytical tools, clinical implementation requires further validation with expanded, heterogeneous datasets to improve robustness and accuracy.

Clinical significance: Machine-learning models that predict the risk of posterior restoration failure-using routinely collected patient, tooth, and diagnostic data-may help dentists tailor recall intervals, prioritize preventive or reparative care, and allocate chair time more efficiently.