Machine Learning-Based Risk Stratification for Survey-Derived Incident Dementia Among Older Korean Adults: A National Longitudinal Person-Period Study

Background/Objectives: Population-based risk stratification for dementia is increasingly important in rapidly ageing societies. Developing prediction models for incident dementia using longitudinal survey data remains challenging because new events are uncommon, observations are repeatedly collected from the same individuals, and predictor variables span multiple health and social domains. This study aimed to develop and compare machine learning models for predicting the two-year risk of newly meeting a survey-derived dementia classification among community-dwelling older adults using longitudinal data from the Korean Longitudinal Study of Aging. Methods: We analysed 12,025 person-period observations from 6380 participants across the 2018–2020 and 2020–2022 intervals after explicitly restricting the risk set to records with known baseline cognition, no baseline dementia classification, and observed follow-up cognition. Sixty baseline predictors were evaluated with eight algorithms using 5-fold participant-level grouped cross-validation and temporally separated validation trained on 2018–2020 and tested on 2020–2022. Performance was assessed using the area under the receiver operating characteristic curve (ROC-AUC), the area under the precision-recall curve (AUPRC), Brier score, calibration, participant-cluster bootstrap confidence intervals (CIs), and top 5%, 10%, and 20% risk-group yield. Results: The survey-derived dementia classification was newly met in 126 of 12,025 person-period observations (1.05%). In grouped cross-validation, CatBoost achieved ROC-AUC 0.890 (95% CI, 0.869–0.910) and AUPRC 0.119 (95% CI, 0.081–0.179); XGBoost achieved ROC-AUC 0.877 and AUPRC 0.113. In temporally separated validation, CatBoost achieved the highest ROC-AUC (0.874); AUPRC values were uniformly low (0.048–0.087) with widely overlapping confidence intervals, so no algorithm was superior for precision-recall performance. The top 10% CatBoost risk group contained 60.3% of incident events in grouped cross-validation and 56.4% in temporal validation, with approximately 6-fold and 5.6-fold enrichment, respectively. Age remained the dominant predictor. Baseline mild cognitive impairment (MCI) was uncommon (39 observations) but strongly enriched for subsequent classification (12 events; 30.8%). Conclusions: Routinely collected longitudinal survey variables supported useful population-level risk ranking for subsequent transition to a survey-derived dementia classification, but absolute case-finding performance remained modest in this rare-event setting. The results support prioritisation for follow-up rather than diagnosis. Sensitivity analyses showed that performance was essentially unchanged after removing baseline survey year or highly missing predictors, whereas excluding baseline MCI reduced precision-recall performance. External validation, recalibration, and prospective evaluation remain necessary before implementation.