Clinical-imaging model for predicting prognosis in contemporary endodontic microsurgery: a retrospective machine learning-based study.

INTRODUCTION: Predictive tools for endodontic microsurgery (EMS) outcomes remain limited. This study evaluated the performance of various machine learning (ML) algorithms in forecasting EMS prognosis using patient-, tooth-, and procedure-related variables. METHODS: A retrospective analysis was conducted on 213 teeth from 180 patients. Clinical and tomographic data were dichotomized and processed using synthetic minority oversampling (SMOTE) to address class imbalance. Feature selection used SelectKbest, chi-square, mutual information, and ensemble classifiers. Several classifiers including logistic regression, random forest, support vector machine, k-nearest neighbors, simple decision tree, and naïve Bayes were trained and validated on an 80:20 split, with performance assessed via accuracy, sensitivity, specificity, precision, F1-score, and area under the receiver operating characteristic curve (AUC). To interpret the model and assess feature importance, the SHapley Additive exPlanations (SHAP) technique was applied. RESULTS: The random forest classifier achieved the highest predictive performance (accuracy: 91%, sensitivity: 91%, specificity: 85%, AUC: 0.97). Eight key predictors of poor prognosis were identified: lack of guided tissue regeneration techniques, poor root-end filling (REF) quality, use of rotary osteotomy, lesion size ≤6.29 mm, patient age >52.50 years, poor root-end resection (RER) quality, steep RER bevel, and suboptimal coronal restoration. CONCLUSION: This study demonstrates that the random forest model showed strong internal performance, but results may be optimistic given the small, SMOTE-augmented dataset and single train-test split. SHAP-derived predictors are clinically plausible yet represent model associations, underscoring the need for external validation before drawing firm clinical conclusions.