Tooth-level prediction of periodontitis progression from baseline clinical characteristics: a retrospective cohort study using Cox regression and SHAP analysis

Tooth-level prognosis can help characterize periodontal risk heterogeneity, yet prediction should be distinguished from treatment effectiveness. Baseline clinical characteristics were evaluated for predicting a composite periodontitis progression endpoint, and Shapley Additive exPlanations (SHAP) were used to interpret model predictions. This single-center retrospective cohort comprised 3,017 teeth from 162 patients. Patients were partitioned at the patient level, with 130 allocated to training and 32 to an internal test set. A single baseline Cox model was used to estimate absolute risks at 1, 3, and 5 years. Patient-cluster robust inference and 500 test-patient bootstrap resamples were applied. Performance was assessed using inverse probability of censoring weighted discrimination, Brier scores, and grouped calibration. SHAP values were used to quantify each variable’s contribution to predicted risk, supporting global interpretation across teeth and tooth-level explanation of model predictions. A total of 420 first composite events were recorded. Internal test AUCs were 0.944 (95% CI 0.852–0.990), 0.912 (95% CI 0.841–0.957), and 0.907 (95% CI 0.862–0.945) at 1, 3, and 5 years, respectively; Brier scores were 0.022, 0.048, and 0.077. Baseline CAL and furcation defect were associated with progression, and furcation defect showed the largest mean absolute SHAP contribution. Removal of baseline severity indicators reduced the 5-year AUC to 0.782. Proportional hazards diagnostics indicated departures from time-constant hazard ratios. Baseline periodontal characteristics provided prognostic information for subsequent tooth-level composite progression in this retrospective cohort. Furcation defect was a prominent model feature; however, interpretation is constrained by measurement uncertainty, predictor-outcome overlap, model assumptions, and limited internal validation. The findings do not establish causal effects or the benefit of specific interventions. External and prospective validation are required before clinical implementation.