Oral Health Indicators as Important Predictors of Frailty in Korean Adults Aged 50 Years and Older: A Sex-Stratified Explainable Machine Learning Approach

Background Frailty is a major public health concern, yet oral health indicators are rarely incorporated into prediction models. Objective To develop sex-stratified explainable machine-learning models for frailty prediction. Methods Cross-sectional data from 12,302 adults aged ≥50 years (KNHANES 2016–2019) were analyzed. Frailty was defined using a modified Fried phenotype, with robust and prefrail participants combined as non-frail. Five algorithms were developed separately for men and women and evaluated on independent test sets. Top-performing models were interpreted using SHAP. Results The survey-weighted frailty prevalence was 14.1% (95% CI, 13.3–14.9). Test-set AUROCs ranged from 0.73 to 0.78 in men and 0.75 to 0.81 in women. XGBoost was selected for interpretation. Chewing and speaking difficulties ranked above hypertension in both sexes. Chewing difficulty showed complementary support in removal analyses. Smoking status and toothbrushing after meals ranked more prominently in men, whereas unmet dental care need and mouthwash use ranked more prominently in women.