Walking Time Emerges as Early Warning Signal for Combined Cognitive and Physical Decline in Older Adults

Older adults whose walking time deteriorates year after year may be showing one of the earliest measurable signs of a dangerous double process: the simultaneous decline of both cognition and physical function. That is the central finding of a new multinational cohort study published in the European Review of Aging and Physical Activity, which used temporal deep learning and explainable artificial intelligence to sift through long-running health data from thousands of older adults. The research, led by Zhenhao Lin and colleagues, set out to identify risk factors for what researchers call dual cognitive-physical decline and to evaluate how physical activity shapes functional trajectories over time.

The team drew on two of the largest longitudinal aging studies in the world. The Health and Retirement Study, covering 11,574 participants, served as the discovery dataset, while the English Longitudinal Study of Ageing, with 3,810 participants, was reserved for external validation. At the heart of the analysis sat a Temporal Attention Network, a deep-learning architecture designed to work with repeated measurements over time, which screened fifteen core variables. The researchers compared this network against six conventional machine-learning algorithms to see which could best anticipate who would slide into combined decline.

When the results came in, a familiar workhorse of applied machine learning took the top spot. XGBoost achieved an area under the curve of 0.898 in the discovery dataset and 0.818 in the external validation cohort, indicating strong discriminatory performance across two different national populations. To make the model’s internal logic transparent, the team applied SHAP, an explainability technique that quantifies how much each variable pushes an individual prediction up or down. The verdict was striking: walking time was the single strongest predictor, carrying a hazard ratio of 1.19. Even more telling, the annual degradation rate of walking time in the dual-decline group was roughly 4.9 times that of healthy controls, suggesting that the pace of deterioration, not just its presence, carries information about looming combined decline.

Physical activity told a consistent protective story across the analysis. All intensity levels of activity were associated with lower hazards, with hazard ratios ranging from 0.92 to 0.95. The researchers also examined transitions between health states using multistate Markov models, which track how people move between categories such as healthy, cognitively impaired, physically impaired, or dually declined over successive study waves. These analyses suggested dynamic cognitive-physical state transitions, with higher physical activity associated with fewer adverse transition patterns.

One of the most concrete illustrations came from the English cohort. Among participants who were healthy at baseline, the probability of transitioning to dual decline by Wave 6 was lower in the high physical activity group, at 1.0 percent with a Wilson 95 percent confidence interval of 0.6 to 1.8 percent, than in the low activity group, at 4.3 percent with a Wilson 95 percent confidence interval of 1.2 to 14.2 percent. The authors themselves urge caution here: the low-activity subgroup was small, which widens the uncertainty around that estimate considerably. The overlapping confidence intervals mean the difference should be read as suggestive rather than definitive.

The study’s conclusions are deliberately measured. Longitudinal deterioration in walking time, the authors write, may serve as a robust early warning marker for dual decline, and physical activity was associated with lower risk and more favorable transition patterns. But they add an important caveat: transitions to less impaired states in the multistate models should not be interpreted as definitive functional recovery. Because the work is observational, the associations between activity and decline cannot be read as proof of causal protection, and the published version is an accepted manuscript shared early, subject to further edits before the final version of record.

Subject of Research: Predicting dual cognitive-physical decline in older adults using physical activity, walking time and deep learning

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

Article References: Lin, Z., Liu, H., Sim, Y.-J., Li, D., Sun, Z., Zhang, J., ShangGuan, Y., & Yan, L. (2026). Predictive value of physical activity and walking time for dual cognitive-physical decline: a multinational cohort study using temporal deep learning and explainable AI. European Review of Aging and Physical Activity. https://doi.org/10.1186/s11556-026-00434-w

Image Credits: AI Generated

DOI: 10.1186/s11556-026-00434-w

Keywords: dual decline, physical activity, walking time, older adults, deep learning, XGBoost, SHAP, HRS, ELSA, cohort study, multistate model, cognitive decline