Data-Driven Tooth Selection Enhances Partial-Mouth Periodontal Screening

The high prevalence of periodontitis has imposed a significant global disease burden. Epidemiologic surveys rely on full-mouth periodontal examination (FMPE) or partial-mouth periodontal examination (PMPE). While FMPE is resource-intensive, the efficiency of PMPE remains questionable. Moreover, the value of subpopulation-tailored PMPE protocols is still unassessed. To address these gaps, we developed an interpretable, data-driven framework that ranks the importance of teeth using SHapley Additive exPlanations (SHAP) values from machine learning models. Using data from the National Health and Nutrition Examination Survey 2009-2014, XGBoost (XGB) and LightGBM (LGB) were trained on maximum interproximal probing depth and clinical attachment loss across 28 teeth from adults aged 35 y and older. Absolute SHAP values were aggregated separately for each model to calculate global tooth importance. The top 10 teeth were then evaluated and benchmarked against the Community Periodontal Index (CPI) and modified Ramfjord protocol. Primary outcomes included quadratic weighted kappa (QWK) for diagnostic agreement with FMPE and the inflation factor (IF) for prevalence estimation bias. The XGB-derived protocol (FDI 47, 27, 17, 26, 37, 16, 42, 46, 36, 41) achieved QWK = 0.85 and IF = 127.29% in the external validation set, with significantly lower IF than CPI (P = 0.003), indicating its potential to replace CPI. Further optimization for age- and gender-specific protocols occasionally yielded localized improvements, such as males aged 55 to 64 y (teeth 17, 16, 27, 47, 26, 33, 45, 37, 43, 25), while the unified XGB-derived protocol remained robust across all strata. These findings support the general use of a unified, data-driven protocol for large-scale epidemiological surveys. SHAP-guided tooth selection offers an interpretable and efficient alternative to traditional PMPE, bridging the gap between accuracy and feasibility in large-scale periodontal surveys.