Joint longitudinal trajectories of physical activity and muscle strength and incident cognitive impairment: evidence from three national aging cohorts

Background Physical activity (PA) and muscle strength are important modifiable factors for healthy aging, yet their joint longitudinal associations with cognitive impairment (CI) remain unclear. We aimed to examine the association between joint longitudinal PA–muscle strength trajectories and incident CI across three national cohorts, and to explore machine learning-based prediction of incident CI. Methods Harmonized data from 11,381 cognitively and functionally intact adults (age ≥ 50 years) from the Health and Retirement Study (HRS), the China Health and Retirement Longitudinal Study (CHARLS), and the Korean Longitudinal Study of Aging (KLOSA) were analyzed. To minimize reverse causality, a landmark design was employed: participants with activities of daily living (ADL) disability at the first trajectory wave or CI at any of the three trajectory-modeling waves were excluded. Group-based multi-trajectory modeling (GBMTM) was used to characterize joint PA–grip strength patterns over these three waves. Subsequent incident CI was tracked after the third wave. Associations were evaluated using inverse probability of treatment weighting (IPTW)-adjusted Cox models, Fine–Gray competing risk models, and discrete-time survival analyses. Stacking ensemble machine learning models with SHAP (SHapley Additive exPlanations) were developed as complementary predictive tools. Results During 6–10 years of post-landmark follow-up, 4,068 participants developed CI. Compared with the “High Activity–High Strength” trajectory, the “Low Activity–Low Strength” trajectory was consistently associated with a higher risk of incident CI across all cohorts: HRS (IPTW-adjusted HR = 1.41, 95% CI: 1.20–1.66), CHARLS (HR = 1.26, 95% CI: 1.03–1.53), and KLOSA (HR = 1.22, 95% CI: 1.02–1.46). Associations for the Low Activity–Low Strength trajectory were directionally similar in Fine–Gray and discrete-time analyses, although some secondary estimates differed across models and subgroups. In exploratory predictive analyses, stacking models generally outperformed individual algorithms within the held-out validation sets, although discrimination remained moderate and was lower than in the training data, and SHAP analyses identified the joint PA–muscle strength trajectory as an important predictive feature in the exploratory models. Conclusions Persistently low PA combined with weak muscle strength was associated with a higher incidence of CI across the three cohorts. These findings support further investigation of joint behavioral–functional trajectories as potential markers of cognitive risk in aging populations.