Predictive performance of gradient boosting, random forest, and logistic regression for periodontitis in adults in the United States

Introduction: Periodontitis is a common chronic condition whose early identification supports prevention, yet it is unclear whether tree ensemble models outperform logistic regression for predicting it from routine sociodemographic and systemic data. Methods: We analyzed adults aged 30 years and older with a complete full-mouth periodontal examination in three cycles of the National Health and Nutrition Examination Survey (NHANES, 2009-2014). Total periodontitis was defined with the CDC/AAP case definitions. Logistic regression, random forest, and gradient boosting were trained on nine sociodemographic and systemic predictors and compared with stratified 5-fold cross-validation using the area under the receiver operating characteristic curve (AUROC) and pairwise DeLong tests. Results: The analytic sample comprised 10,642 adults, of whom 4,903 (46.07%) had periodontitis. Gradient boosting achieved the highest discrimination (AUROC 0.71; 95% CI 0.71 to 0.72), significantly higher than logistic regression (AUROC 0.70) and random forest (AUROC 0.68). Random forest did not outperform logistic regression. Age, poverty-income ratio, sex, race/ethnicity, and body mass index were the most influential predictors. Conclusions: Gradient boosting predicts periodontitis more accurately than logistic regression or random forest using readily available variables, whereas random forest offers no advantage over logistic regression.