Longevity & AgingResearch PaperOpen Access

Wearables and AI show early promise for spotting cognitive decline, but validation lags

A systematic review of 49 studies links disrupted sleep and rest-activity rhythms to worse cognition. AI models reach AUCs of 0.70-0.95 but lack external validation.

Friday, November 20, 2026 0 views
Published in J Med Internet Res
Older adult asleep wearing a smartwatch, with glowing rest-activity rhythm waves and a faint brain outline above

Summary

This systematic review pooled 49 studies (2020-2025) on wearable-derived data for early detection of cognitive impairment and dementia. Most used research-grade actigraphy, and most analyses were statistical. Fragmented sleep, weakened circadian rhythms and irregular activity were linked to worse cognition, with modest-to-moderate effect sizes. Machine and deep learning models reported AUCs of about 0.70-0.95. Only about a quarter of studies directly addressed early detection or prevention through risk estimation or prediction. Samples were often small, monitoring was short, and external validation was limited. The evidence is associational and mostly diagnostic or risk-stratification, not proof that wearables improve outcomes. The review was not preregistered and did not meta-analyze the results.

Detailed Summary

Standard cognitive tests such as the MMSE and MoCA are episodic, so they can miss gradual changes that precede dementia. Wearables can passively record sleep, activity and circadian rhythms over long periods, and AI could turn those streams into digital biomarkers for earlier risk detection. This review asked how well the evidence supports that idea.

The authors searched PubMed, Scopus, ACM Digital Library and SpringerLink for peer-reviewed human studies published January 2020 to December 2025. Participants had a mean age of 50 or older, wearable data covered at least 24 hours, and cognitive outcomes were validated. Two reviewers screened studies and assessed bias. The review was not prospectively registered. Because devices, outcomes and methods varied so much, there was no meta-analysis, only a narrative synthesis.

Forty-nine studies were included, with sample sizes from 14 to 91,948 (over 200,000 in total, median 145). Research-grade actigraphy dominated (43/49), and commercial devices appeared in 7. The MMSE was the most common outcome (18/49). Most analyses were statistical (36/49), with 7 using machine learning and 6 using deep learning. Design was split between cross-sectional (21) and longitudinal cohort (24) studies. Most studies were rated low risk of bias (about 69-71%, with the paper reporting slightly different figures).

Statistical analyses tied disrupted sleep, fragmented circadian rhythms and irregular activity to worse cognition, with modest-to-moderate effect sizes. Machine and deep learning classifiers reported AUCs of roughly 0.70-0.95. About a quarter of studies (13/49) addressed early detection or prevention through longitudinal risk estimation or predictive modeling. The rest were indirect correlational findings.

The signal is a behavioral marker of risk, and it works at the level of detection and stratification. The review does not show that wearable-guided action changes cognitive outcomes or dementia incidence. Caveats include small samples, short monitoring, limited external validation, and wide AUC ranges that likely reflect weaker designs at the low end. Large cohorts (47,000-92,000 participants) weighed heavily on aggregate figures. Reverse causation is also possible, since early neurodegeneration can itself disrupt sleep and rhythms. Wearable biomarkers look worth watching, but they are not yet validated for individual screening.

Key Findings

  • 49 studies (2020-2025) included; 87.8% used research-grade actigraphy and 14.3% used commercial wearables.
  • Disrupted sleep, circadian fragmentation and irregular activity were associated with worse cognition, with modest-to-moderate effect sizes.
  • Machine and deep learning classifiers reported AUCs of about 0.70-0.95; external validation was limited.
  • Only 13/49 studies (26.5%) addressed early detection or prevention via longitudinal risk estimation or prediction.
  • No meta-analysis was possible; the review was not preregistered, and evidence of improved outcomes from wearable-guided action is absent.

Methodology

Systematic review following PRISMA 2020 across four databases, with dual independent screening and risk-of-bias assessment (ATCS, Newcastle-Ottawa, Cochrane, QUADAS-2). Narrative synthesis only, because of heterogeneity; not preregistered. Sample sizes ranged from 14 to 91,948.

Study Limitations

Included studies often had small samples, short monitoring and little external validation, and associations may reflect reverse causation. Heterogeneity prevented pooling, the review was unregistered, and the search was English-only with 15 reports unretrievable. Reported AUCs span a wide range and may be optimistic given limited external validation.

Enjoyed this summary?

Get the latest longevity research delivered to your inbox every week.

Enter your email to subscribe: