Predictive factors for dementia among older adults in South Korea: an interpretable machine learning analysis

South Korea is among the fastest-aging countries globally, with a rapidly rising prevalence of dementia. Early identification of individuals at risk is critical for effective prevention, as dementia is influenced by both non-modifiable factors, such as age, sex, and baseline cognitive status, and modifiable factors, including socioeconomic conditions, health behaviors, and psychosocial characteristics. This study aimed to identify multidimensional determinants of dementia using machine learning applied to nationally representative longitudinal data, examining how these factors interact across demographic and cognitive subgroups to inform targeted, evidence-based prevention strategies.

We analyzed data from the Korean Longitudinal Study of Aging (KLoSA; 2014-2020), including 4,958 participants aged 45 years and older without baseline dementia. Participants were stratified by baseline cognitive status (cognitively normal vs. mild cognitive impairment (MCI)), with further subgroup comparisons by age (< 65 vs. ≥ 65) and sex for cognitively normal individuals. Predictors spanning sociodemographic, health, behavioral, and contextual domains were examined. Four regression algorithms-linear regression, random forests, XGBoost, and CatBoost-were applied, and model performance was evaluated via RMSE, MAE, and R². Predictor importance was assessed using a multi-method approach integrating model-based metrics and SHAP values, with top predictors identified for each subgroup.

Predictive performance was comparable across algorithms, with R² ranging from 0.201 to 0.361, highest in the MCI_All dataset. Age and education were consistently the most influential non-modifiable factors. Key modifiable contributors included oral health, depression, household income, quality of life, and IADL performance. Importance patterns varied by cognitive status, age, and sex: socioeconomic and psychosocial factors were more influential in cognitively normal adults, whereas health status and IADL predominated in MCI participants. Age-stratified analyses highlighted oral health, depression change, and social contact in adults < 65, and cumulative factors including IADL decline in adults ≥ 65. Sex-stratified analyses indicated stronger effects of household income and social engagement in men, and depression and oral health in women. SHAP analyses confirmed inverse associations between changes in depression and IADL performance and predicted cognitive scores.

Age and education were the strongest predictors of cognitive function, while modifiable factors-including oral health, depression, social engagement, and IADL performance-played significant roles across subgroups. This interpretable machine learning approach revealed nuanced patterns of predictor importance across cognitive status, age, and sex, underscoring the value of targeted interventions to reduce dementia risk in aging populations.