Machine learning prediction of periodontitis across age strata: using data from the Korea National Health and Nutrition Examintaion Survey

Abstract

Objectives: Most existing machine learning (ML) models are trained on mixed-age samples and rarely examine whether performance varies across life stages.We compared age-stratified ML models for predicting periodontitis in a nationally representative sample of Korean adult population.Methods: We trained five ML models using 16 features from the seventh Korea National Health and Nutrition Examination Survey (KNHANES Ⅶ) to predict periodontitis according to four age strata: adults ≥19 years (n=12,689), 19-44 years (n=4,847), 45-64 years (n=4,861), and ≥65 years (n=2,981).The performance of all models was evaluated with stratified 5-fold cross-validation, combined with class-weight correction to address the imbalance in periodontitis prevalence, and 95% confidence intervals (CI) estimated using 1,000 bootstraps resamples.Results: Predictive performance declined with advancing age: the best-performing model achieved an AUC of 0.763 (95% CI, 0.745-0.782) in adults aged 19-44 years, 0.684 (0.668-0.699) in adults aged 45-64 years, and only 0.567 (0.547-0.587) in adults aged ≥65 years, corresponding to an absolute AUC reduction of approximately 0.196 between the youngest and oldest strata.This agerelated degradation was observed uniformly across all five algorithms.Moreover, in the ≥65 years stratum none of the tested algorithms reached clinically useful discrimination (all AUCs <0.60 with upper 95% confidence limits below 0.60).Conclusions: These findings demonstrate that the predictive value of survey-based features for periodontitis is strongly age-dependent.Therefore, ML-based periodontal screening should be designed and validated separately for each age stratum.