Performance of a Chinese Cognitive Decline Risk Model in a Japanese Cohort: A Validation Study

ObjectivesTo develop a simple risk prediction model for cognitive decline in a Chinese older adult cohort, and to evaluate its performance and transportability through temporal validation and external validation in a Japanese older adult cohort. MethodsThe prediction model was developed using a derivation cohort of 5,985 cognitively normal older adults from the China Health and Retirement Longitudinal Study (CHARLS, 2011-2015). A comparison of seven machine learning algorithms was conducted, and the standard Cox Proportional Hazards (CoxPH) model was selected based on its optimal balance of performance and parsimony. The final model was then validated on a temporal cohort (CHARLS 2015-2018, n=1,333) and an external cohort (Japanese Study of Aging and Retirement [JSTAR] 2007-2009, n=2,798). A comprehensive preprocessing pipeline, including Iterative Imputation for high-missingness predictor variables and One-Hot Encoding for categorical variables, was developed on the training data and applied to all cohorts. Model performance was assessed via discrimination, calibration, risk stratification and clinical utility. ResultsIn temporal validation, the model demonstrated strong performance with an AUC of 0.72 and reliable calibration (Slope = 1.02). In the external JSTAR cohort, the model maintained high discriminative power (AUC = 0.68), which was even superior to the development set (AUC = 0.62). However, a notable calibration shift was observed (Slope = 1.54), indicating a systematic underestimation of absolute risk in the low-prevalence Japanese population. While decision curve analysis (DCA) showed substantial net benefit in the temporal cohort, its utility in the external cohort was most effective within a narrow threshold range near the population prevalence. Sensitivity analyses confirmed that the models risk-ranking ability remained robust across 2-year and 4-year horizons. ConclusionOur 6-predictor model shows robust risk-ranking consistency across cohorts, but absolute risk estimates are sensitive to population and temporal differences. While effective for identifying high-risk individuals, local recalibration is essential for accurate clinical prognosis in new settings.