Wrist wearables show no link between sleep metrics and anxiety in a 44-study meta-analysis
A meta-analysis of 44 wearable studies found no sleep-anxiety link, while low activity and high heart rate loosely tracked anxiety. Wearables work best combined with other data.
Summary
Researchers pooled 44 studies (42 articles) testing whether data from wrist-worn wearables track anxiety in adults. Sample sizes ranged from 17 to 170,320. Pooled analyses of four sleep measures (sleep efficiency, time awake after sleep onset, total sleep time, sleep onset latency) found no significant association with anxiety. Narrative syntheses, based on fewer studies, suggested lower physical activity and higher heart rate go with higher anxiety symptoms. Machine learning models using wearable data alone performed unevenly, and improved when combined with self-report or clinical data. The inferential studies were mostly of fair quality and observational (cross-sectional or longitudinal), so they cannot show cause. Wearables look more promising as one input among several than as stand-alone anxiety screeners.
Detailed Summary
Anxiety is the most common mental health condition worldwide, affecting an estimated 301 million people in 2019, and in older adults it has been linked to higher dementia risk. Current assessment relies on self-report questionnaires and clinical interviews, which are retrospective, resource-heavy and offer only snapshots. Passive, continuous data from wrist-worn wearables could in principle fill that gap, so this review asked whether wearable-derived digital biomarkers are associated with anxiety.
The authors searched six databases, including gray literature, through September 2025 (PROSPERO CRD42023409995). They included cross-sectional or longitudinal studies of adults using research- or consumer-grade wrist devices, with anxiety measured by standardized assessments (self-report questionnaires or clinical diagnoses). Forty-four studies from 42 articles qualified: 36 used inferential statistics (21 sleep, 8 physical activity, 2 heart rate variability, 5 multiple types) and 8 used machine learning. Sample sizes ranged from 17 to 170,320.
The four sleep meta-analyses were all null. Sleep efficiency gave Fisher z=-0.07 (95% CI -0.14 to 0.002; P=.06), wake after sleep onset z=0.13 (CI -0.04 to 0.30; P=.11), total sleep time z=0.009 (CI -0.01 to 0.03; P=.28), and sleep onset latency z=0.04 (CI -0.07 to 0.15; P=.08). Prediction intervals were wide (for example -0.19 to 0.05 for sleep efficiency), meaning a future study could plausibly find a modest effect in either direction. Random-effects models were used, with sensitivity analyses by anxiety type (state or trait) and age group.
Physical activity and heart rate could not be pooled and were synthesized narratively. Lower physical activity and higher heart rate were associated with greater anxiety symptoms, though this rests on limited studies. The eight machine learning studies showed varied performance when using wearable data alone, and prediction improved when wearable data were combined with other sources such as self-report and clinical data.
The practical message is that a null on four sleep metrics bounds the claim that actigraphy-style sleep measures track anxiety, but it does not dismiss the activity and heart rate signals, which remain plausible and worth testing in better designs. Authors conclude digital biomarkers are likely more useful when integrated with other information than used alone as screening tools.
Caveats are substantial. Most inferential studies were rated fair quality, the designs were observational (cross-sectional or longitudinal), and sensitivity analyses had to account for state versus trait anxiety. Studies that could not be pooled were synthesized narratively, and association does not establish that activity or heart rate changes cause or predict anxiety. There is no validated quality assessment tool for machine learning studies, so the authors adapted a modified QUADAS-2.
Key Findings
- 44 studies from 42 articles met criteria: 36 inferential (21 sleep, 8 activity, 2 HRV, 5 multiple) and 8 machine learning; N ranged 17 to 170,320
- Sleep efficiency was not significantly associated with anxiety (Fisher z=-0.07, 95% CI -0.14 to 0.002, P=.06)
- Wake after sleep onset (z=0.13, P=.11), total sleep time (z=0.009, P=.28) and sleep onset latency (z=0.04, P=.08) were all null
- Narratively, lower physical activity and higher heart rate were associated with greater anxiety symptoms, based on limited studies and not pooled
- Machine learning on wearable data alone performed variably; performance improved when combined with self-report or clinical data
- Inferential studies were mostly of fair quality on the NHLBI tool; designs were cross-sectional or longitudinal
- Prior machine learning meta-analyses reported about 82% accuracy for anxiety classification, cited as context
Methodology
Systematic review and meta-analysis following PRISMA, registered in PROSPERO, searching 5 databases plus ProQuest and Google Scholar through September 2025, with dual independent screening (interrater agreement assessed by Cohen kappa). Risk of bias was assessed with the NHLBI tool for inferential studies and a modified QUADAS-2 for machine learning studies. Random-effects meta-analyses were run per digital biomarker where possible, with prediction intervals reported and sensitivity analyses by anxiety type (state or trait) and age group. Activity, heart rate and machine learning studies were synthesized narratively.
Study Limitations
Included inferential studies were mostly of fair quality and observational (cross-sectional or longitudinal), limiting causal inference. Studies that could not be pooled were synthesized narratively, and the activity, heart rate and machine learning findings rest on limited studies. No validated quality tool exists for machine learning studies, so a modified QUADAS-2 was used. No conflicts of interest were reported in the supplied text.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe:
