A parsimonious six-predictor model for memory decline: cross-population validation in Chinese and Japanese aging cohorts.

To develop and validate a parsimonious risk model for short-term memory decline in older adults and to evaluate its cross-population transportability between Chinese and Japanese cohorts. The model was developed in 5985 cognitively normal older adults from the China Health and Retirement Longitudinal Study (CHARLS, 2011-2015). Seven machine learning algorithms were compared, and a Cox proportional hazards (CoxPH) model was selected for its optimal balance between performance and parsimony. The final model was validated in a temporal CHARLS cohort (2015-2018; n = 1333) and an external Japanese cohort from the Japanese Study of Aging and Retirement (JSTAR, 2007-2009; n = 2798). Performance was assessed using discrimination, calibration, decision curve analysis, and bootstrap-derived confidence intervals. In temporal validation, the model demonstrated good discrimination (C-index = 0.72) with acceptable calibration (slope = 1.40). In the external JSTAR cohort, discriminative performance remained moderate and stable (C-index = 0.68), and calibration was comparable (slope = 0.96) despite differences in baseline incidence and follow-up duration. Decision curve analysis showed net benefit in the temporal cohort and consistent risk stratification in the external cohort. Sensitivity analyses confirmed stable performance across varying follow-up horizons. The six-predictor model consistently stratified short-term memory decline risk across distinct East Asian populations. The findings support its cross-population transportability for relative risk stratification in aging cohorts.