An interpretable postoperative model predicts one-year healing after endodontic microsurgery with temporal validation and clinician benchmarking

Endodontic microsurgery is an important tooth-preserving treatment for periapical lesions, but accurate postoperative prognosis remains challenging. This study aimed to develop and temporally validate an interpretable model for one-year healing after endodontic microsurgery and to benchmark its discrimination against clinician estimates. The development cohort comprised 312 patients (273 healed and 39 failed) and was split into a training set ( n = 218; 191 healed and 27 failed) and an internal model-selection set ( n = 94; 82 healed and 12 failed). The temporal cohort comprised 108 patients (95 healed and 13 failed). Twenty-six candidate predictors were evaluated using least absolute shrinkage and selection operator and Boruta procedures within the training set, and six machine-learning models were compared. The final L2-penalized logistic regression model was interpreted using exact linear SHapley Additive exPlanations (SHAP), exploratory predictor-effect profiles, and a separate post-hoc odds-ratio refit, and was benchmarked against probability estimates from five endodontic clinicians. Logistic regression achieved an out-of-fold training AUC of 0.812 (95% CI, 0.708–0.907), an internal model-selection AUC of 0.868 (95% CI, 0.761–0.949), and a temporal-validation AUC of 0.703 (95% CI, 0.530–0.867). The internal model-selection Brier score was 0.084 (95% CI, 0.044–0.129). Internal model-selection calibration intercept and slope were 0.80 (95% CI, − 0.36 to 1.91) and 0.60 (95% CI, 0.24 to 1.39); temporal estimates were 1.11 (95% CI, 0.32 to 2.05) and 0.54 (95% CI, 0.16 to 1.15). Buccal bone plate height and CBCT bone-loss volume were the dominant SHAP contributors. The paired clinician benchmark analysis yielded model and clinician AUCs of 0.696 and 0.771, respectively, with an AUC difference of − 0.075 (DeLong P = 0.186). This study presents an interpretable postoperative/perioperative prognostic model for one-year healing after endodontic microsurgery. Temporal performance was lower than the internal model-selection estimate, and the five-clinician benchmark provided a comparison with clinical estimates. Multicenter validation and development of a fully preoperative model are priorities for future work.