Development and temporal validation of a machine-learning based risk prediction model for depression in older adults...

Depression in older adults is a multifactorial condition influenced by demographic, behavioral, and health-related factors. Oral functional problems, such as chewing difficulty, have received relatively limited attention despite potential relevance to mental health. Machine-learning models were developed and temporally validated to predict depressive symptoms while assessing the importance of chewing difficulty alongside other health predictors. Although these approaches enable the evaluation of multiple competing variables, limitations remain in capturing complex interactions among demographic, behavioral, and clinical factors within large population datasets.

This prediction modeling study used KNHANES data from Korean adults aged ≥ 65 years who completed the PHQ-9 in 2020, 2022, and 2024. After excluding participants with missing PHQ-9 data and applying complete-case preprocessing across 24 predictors, 3,881 participants were included. The 2020 and 2022 cohorts were used as the development set, and the 2024 cohort was reserved as an independent temporal test set. Five machine-learning models were trained for PHQ-9 ≥ 10 and PHQ-9 ≥ 5 outcomes using stratified 5-fold cross-validation, with preprocessing, class imbalance handling, and hyperparameter tuning performed on the development set. Final tuned models were evaluated on the 2024 test set using area under the receiver operating characteristic curve (AUROC) as the primary metric, with area under the precision-recall curve (AUPRC), sensitivity, specificity, F1 Score, and Brier score also reported.

CatBoost demonstrated the best discrimination for both outcomes. For PHQ-9 ≥ 10, CatBoost achieved AUROC 0.853 (95% CI 0.780-0.911), sensitivity 0.650, specificity 0.877, AUPRC 0.267, and Brier score 0.024. For PHQ-9 ≥ 5, AUROC was 0.751 (95% CI 0.707-0.789), sensitivity 0.620, specificity 0.764, AUPRC 0.300, and Brier score 0.088. Across models, perceived stress and activity limitation were the strongest predictors. Chewing difficulty ranked as the most important oral health predictor (6th and 5th among 24 variables for PHQ-9 ≥ 10 and ≥ 5, respectively) and consistently higher model-predicted probability of depressive symptoms across SHAP analyses.

Machine-learning models demonstrated robust temporal validity for predicting late-life depression. Chewing difficulty showed consistent predictive importance independent of general health factors, supporting the integration of oral functional assessment into multidomain depression risk screening strategies.