Predictive value of physical activity and walking time for dual cognitive-physical decline: a multinational cohort study using temporal deep learning and explainable AI

Abstract

To develop an interpretable deep-learning framework to identify risk factors for dual cognitive-physical decline in older adults and evaluate the impact of physical activity (PA) on functional trajectories. Using HRS ( n  = 11,574) for discovery and ELSA ( n  = 3,810) for external validation, we employed a Temporal Attention Network (TAN) to screen 15 core variables. Six conventional machine-learning algorithms and TAN were compared, with SHAP used to interpret the best-performing model. PA associations were analyzed using Cox regression and multistate Markov models. XGBoost performed best (HRS AUC = 0.898; ELSA AUC = 0.818). Walking time was the top predictor (HR = 1.19), with its annual degradation rate in the dual-decline group being ~ 4.9 times that of healthy controls. All PA intensities were associated with lower hazards (HR: 0.92–0.95). Among participants who were Healthy at baseline in ELSA, the probability of transitioning to Dual Decline by Wave 6 was lower in the High PA group (1.0%, Wilson 95% CI: 0.6–1.8%) than in the Low PA group (4.3%, Wilson 95% CI: 1.2–14.2%). However, this PA-stratified transition estimate should be interpreted cautiously because the Low PA subgroup was small. Multistate Markov analyses suggested dynamic cognitive-physical state transitions, with higher PA associated with fewer adverse transition patterns. Longitudinal deterioration in walking time may serve as a robust early warning marker for dual decline. PA was associated with lower risk and more favorable transition patterns, although transitions to less impaired states should not be interpreted as definitive functional recovery.