Predicting Root Canal Retreatment Outcome Using a Machine Learning and Radiomics Approach

Objective To explore the feasibility of developing an interpretable machine learning (ML) model that integrates clinical variables and quantitative radiomics features to predict 12-month outcomes of nonsurgical root canal retreatment (NSreRCT). Materials and Methods In this retrospective case-control study, 200 teeth (100 failed and 100 successful NSreRCT) were included, matched by age, tooth type, and periapical status. For each case, 30 preoperative clinical variables and periapical radiographs were collected. Radiomics features were extracted from the core and boundary regions of periapical lesions. Seven ML classifiers were trained and validated using the 5/10 most informative radiomics and/or clinical variables, identified through supervised feature selection nested within stratified 5-fold cross-validation. Within each training partition, 448 candidate pipelines were ranked using an internal hold-out evaluation. Model performance was assessed using accuracy, sensitivity, specificity, Matthew's Correlation Coefficient, Area Under Precision-Recall Curve, mean Area Under Receiver Operating Characteristic Curve (AUC) and stacked AUC. Feature importance was evaluated using SHapley Additive exPlanations analysis. Results The Rad10AndCli10 model incorporating the top 10 radiomics features and top 10 clinical variables achieved the highest performance (stacked AUC: 0.902), outperforming the Rad5AndCli5 model (stacked AUC: 0.896) and Rad10 model (AUC: 0.866) or Cli10 model (AUC: 0.750). The Rad10AndCli10 model demonstrated stable performance across all 5 folds (AUC range: 0.862-0.945). Radiomics features from lesion boundaries contributed most to predictions, and higher lesion heterogeneity was associated with a greater risk of treatment failure. Clinical predictors including deepest periodontal probing depth and size of periapical lesion size, were consistently selected. Conclusion ML models combining radiomics features and clinical variables improved prediction of NSreRCT outcomes compared with models using either data source alone. Radiomics features, particularly those derived from periapical lesion boundaries, contributed substantially to prediction, and greater lesion heterogeneity was associated with a higher risk of NSreRCT failure.