Machine learning-based prediction model for 2-year tooth loss risk in periodontitis patients: integrating radiographic features and clinical parameters

To develop and internally–externally validate a multidomain machine-learning (ML) model for 2-year tooth loss in periodontitis patients using exclusively baseline predictors, and to quantify the incremental value of radiographic over clinical features. This multicenter retrospective cohort included 862 periodontitis patients (Stage I–IV) from three centers (2019–2024), of whom 718 were evaluable at a prespecified 24-month prediction horizon. Forty-one baseline predictors across six domains plus three a priori interaction terms were evaluated using seven ML algorithms with stratified 5-fold cross-validation, with all preprocessing (multiple imputation, standardization, feature engineering) performed within each training fold. Generalizability was assessed via leave-one-center-out (LOCO) internal–external and temporal validation. Incremental radiographic value was quantified through nested logistic regression with category-free NRI/IDI. Ensemble-level SHAP provided interpretability; DCA assessed clinical utility; Cox regression analyzed time to tooth loss. The 2-year event rate was 29.0% (208/718). Random forest achieved the numerically highest discrimination (AUC 0.886, 95% CI 0.855–0.913; Brier 0.112), statistically indistinguishable from the stacking ensemble (AUC 0.881, 95% CI 0.849–0.910; ΔAUC − 0.005, p = 0.198). For the primary stacking ensemble, leave-one-center-out (LOCO) internal–external validation yielded a mean AUC of 0.885 and temporal validation an AUC of 0.866; discrimination was essentially unchanged in landmark analyses excluding events within 3–6 months of baseline (AUC 0.850–0.867). Clinical parameters dominated SHAP importance (46.1% of total attribution after reassigning composite severity scores to their source domains), followed by radiographic features (22.9%). Adding clinical features to the demographic/behavioral baseline produced a substantial gain (ΔAUC + 0.284, p < 0.001), while the radiographic increment was non-significant in nested regression (ΔAUC − 0.003, p = 0.494); radiographic information contributed primarily via nonlinear interactions (Smoking×RBL%). Cox regression confirmed RBL%, CAL_Mean, Smoking, and Diabetes as predictors. DCA showed net benefit at thresholds of approximately 9–80%. A multidomain ML model restricted to baseline predictors and evaluated at a fixed 2-year horizon retained robust discrimination and internal–external generalizability, supporting its further development as a decision-support tool for personalized periodontal risk stratification; independent external validation remains necessary.