This review was conducted in line with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) recommendation [21] and registered with PROSPERO (International Prospective Register of Systematic Reviews) in March 2023 (CRD42023409995). The PRISMA 2020 checklist [22] (Multimedia Appendix 1), the PRISMA Abstract checklist [22] (Multimedia Appendix 2), and the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses–Search Extension) checklist [23] (Multimedia Appendix 3) were used.

A total of 5 databases (Medline, Embase, PsycINFO, Web of Science, and CINAHL) were systematically searched. Further, unpublished gray literature searches were conducted with ProQuest Dissertations and Theses Global. Following published recommendations [24], we also checked the first 300 papers on Google Scholar to supplement the gray literature search. Finally, to supplement the electronic search, references of eligible papers were reviewed to identify any additional eligible studies.

The search strategy combined 2 search term strings with “AND.” Each string reflected a concept related to the research question: (1) digital biomarkers and (2) psychosocial risk factors of dementia (Multimedia Appendix 4). The search strategy was informed by previous reviews in related research areas [17], but all search terms and strategies were developed specifically for this review and adapted for each database. There were no restrictions imposed on the date of publication. This review did not restrict the database searches to human studies and studies published in English, and no filters were used. The initial searches were conducted on April 27, 2023; a subsequent search was done on September 24, 2024, and a final search focusing only on anxiety was done on September 21, 2025. While the PROSPERO registration included multiple psychosocial risk factors, this systematic review focused specifically on anxiety, due to the large number of articles that made it past full-text screening. As all eligible anxiety studies in earlier searches were identified via anxiety-specific terms, a final search restricted to anxiety-related search terms was conducted in 2025.

The web platform Covidence (Veritas Health Innovation, 2020) was used for deduplication and to coordinate multiuser title-abstract and full-text screening. Two reviewers independently screened titles and abstracts of all the identified studies (YL screened 100% of the articles, and NC, HAG, RM, and HD-K collectively screened 100%), followed by full-text screening. Any disagreements were resolved by a third reviewer (NM or HD-K), with HD-K only resolving conflicts for articles she had not screened at the title-abstract screening stage. Interrater agreement was assessed using the Cohen κ coefficient.

In accordance with the PROSPERO registration, our search strategies encompassed a range of psychosocial risk factors for dementia. However, given the large number of eligible articles that made it through to the full-text screening stage, we decided to categorize these articles into different psychosocial risk factors before the data extraction stage.

This systematic review focused on anxiety, both clinical anxiety and anxiety symptoms. The PECOS (Population, Exposure, Control, Outcome, and Study design) framework was used to structure the study selection. Studies were selected if they (1) involved adults with a mean age of >18 years (Population); (2) reported data from research- or consumer-grade wrist-worn devices (Exposure); (3) did not require a control group, although studies including participants without anxiety were considered as comparators when applicable (Control); (4) assessed GAD or anxiety symptoms using standardized assessments (eg, self-report questionnaires or clinical diagnoses [Outcome]); and (5) used cross-sectional or longitudinal designs (Study design). Studies published in an English-language, peer-reviewed journal or thesis/dissertation, and studies that examined the relationship between anxiety and digital biomarkers via inferential statistics or machine learning methods were also eligible. Studies were excluded if they (1) focused on a sample that received diagnoses of neurodegenerative disorders or physical health conditions (however, studies were eligible if participants without these conditions were analyzed separately), and (2) primarily focused on participants with a psychiatric comorbidity, neurological disorder, or stroke (>50% of the sample).

A standardized form was created to extract the following data from eligible inferential statistical studies: (1) authors and year of publication; (2) study characteristics, including type of study (ie, cross-sectional or longitudinal); (3) demographic information; (4) characteristics of digital biomarkers, including digital biomarker type and devices used for data collection; (5) anxiety measurement, and (6) data required for meta-analysis (eg, mean difference and correlation coefficients). For studies using machine learning approaches, another standardized form was created to extract the following details: (1) authors and year of publication, (2) demographic information, (3) ground truth assessment, (4) characteristics of digital biomarkers (5) machine learning task, (6) predictor features, (7) algorithms, (8) validation method, and (9) model performance.

Where necessary, authors were contacted to request additional information.

Effect sizes were combined across studies, separately for each type of digital biomarker of health, using random-effects meta-analyses. When effect sizes were unavailable, other available data were used for calculations (eg, t test statistic). Random-effects meta-analyses were used as they take into account between-study heterogeneity [28]. For meta-analysis based on a limited number of studies (<20 studies [29]), we applied the Hartung-Knapp-Sidik-Jonkman adjustments when calculating CIs and P values, which are recommended for small meta-analyses [30]. Prediction intervals for random-effects meta-analyses were also reported [31].

To ensure only 1 effect size from each included study was used in the primary meta-analysis, studies were selected based on an a priori determined hierarchy. Specifically, we prioritized (1) the largest sample size, (2) unadjusted over adjusted estimates, and (3) trait (chronic/ongoing) over state (present moment) anxiety, when more than 1 study used the same cohort. Trait anxiety refers to the general tendency to respond with anxiety across various situations, whereas state anxiety is a temporary response reflecting the present moment [32]. Since trait anxiety reflects a chronic condition with long-term implications and offers a more stable measure of general anxiety levels, it was the focus of our analysis.

Sensitivity analyses were conducted to examine whether results differed according to anxiety type (trait and state anxiety) and age group (young, middle, and older).

For each meta-analysis, heterogeneity was assessed using τ² (between-study variance) and I² (proportion of total variation in study estimates that is due to heterogeneity [33]). When there were 10 or more studies included in a meta-analysis, funnel plots would be used to assess small-study effects [34].

All analyses were conducted using the metafor package in R software (version 4.2.2; R Foundation for Statistical Computing).

Qualitative synthesis was conducted for studies using inferential statistical approaches that were not included in the meta-analysis. These studies were synthesized based on the type of digital biomarkers of health examined (ie, sleep metrics and activity metrics). On the other hand, studies using machine learning approaches were not meta-analyzed and were synthesized qualitatively, ordered by performance metrics (ie, accuracy, sensitivity, precision, and F₁-score).

After the removal of duplicates, title/abstract screening, and full-text review, 365 articles were eligible (Figure 1). Of these, 47 articles focused on anxiety. Interrater reliability was high, with Cohen κ of 0.73 (substantial agreement) and 0.83 (almost perfect agreement) at the title/abstract screening and full-text review stages, respectively.

Of the 8 studies included in this review, some studies used multiple algorithms within the same study, each with multiple performance metrics. Accuracy measures correct classifications (ie, presence/absence of anxiety). Sensitivity is the ability to detect true positives (ie, correct identification of individuals with anxiety), while precision is the correct classification among predicted positives (ie, how many individuals classified as having anxiety actually have anxiety). F₁-score balances precision and sensitivity, and mean absolute error (MAE) measures the average prediction error, with lower MAE signifying better performance.

A total of 3 studies reported 13 accuracy estimates, ranging from 56.3% to 94.6% [70,71,73]. Two studies reported 9 sensitivity estimates (range 56.3%-94.6%), 9 precision estimates (range 56.3%-94.8%), and 9 F₁-score estimates (range 55.9%-94.6%) [71,73]. Overall, 56 MAE estimates were reported from 2 studies [75,76], ranging from 0.05 to 0.63.

One article using 2 cohorts found better accuracy in predicting anxiety using frequency-domain HRV and nighttime HRV (range 63.9%-69.4%) data than time-domain HRV and daytime (range 63.2%-63.4%) data [70]. They also found that nighttime HRV data (maximum accuracy: 69.4%) performed better than daytime data (maximum accuracy: 63.2%) [70].

One study with older adults who had mild cognitive impairment found that combining activity (24-hour activity rhythm), sleep, and anxiety questionnaire items improved model performance across accuracy (91.0%-94.6% vs 56.3%-59.9%), precision (91.4%-94.8% vs 56.4%-61.0%), recall (91.0%-94.6% vs 56.3%-59.9%), and F₁-scores (90.9%-94.6% vs 55.9%-59.0%) compared to models that excluded anxiety questionnaire data [73]. Similarly, another study using the same sample found that combining activity (step count) and sleep stages achieved a predictive success rate for participants without anxiety (98%) but had a low predictive success rate identifying those with anxiety (18%) [74]. Models including anxiety questionnaire items improved the predictive success rate for identifying participants with anxiety (82%) and maintained a good predictive success rate at identifying those without anxiety (95%). These studies did not report feature importance in their models.

A study with university students used digital biomarkers of health—HRV, electrodermal activity, and skin temperature—collected from wrist-worn wearables and smart shirts to predict low versus high anxiety levels. The best classifier (support vector machine) achieved an overall accuracy of 90.9%, with a precision of 0.90-0.92, recall of 0.90-0.92, and F₁-score of 0.91 [72].

Two studies predicted anxiety in office workers [71,75]. One study, using 13 sleep metrics, achieved an F₁-score of 75.6% and identified that the top 5 important predictors were REM sleep time ratio, REM sleep minutes, light sleep time ratio, deep sleep time ratio, and total wake time ratio [71]. Another study, using data from wearable devices (activity, stress, sleep, and heart rate data) and smartphones, found that the best non–time-based model (ie, without incorporating temporal data) predicted anxiety with an MAE of 0.6316. Time-based models showed improved performance, with an MAE of 0.3729 when using data from the previous 3 days to predict the next day’s anxiety [75]. They found that the top 20 predictors included measures of sleep (bedtime, wake time, and sleep duration), stress (low stress duration), and activity (highly active duration).

Finally, 1 study comparing 4 modeling approaches—conventional, multitask learning approach (model trained to solve multiple tasks simultaneously), time-based approach, and time-based multitask approach—found that time-based models using data from the previous 15 days to predict next day’s anxiety achieved the lowest MAE (0.0047), followed by the time-based multitask model (MAE 0.0083), the conventional model (MAE 0.1277), and the multitask model (MAE 0.1347) [76]. While activity, sleep, and heart rate were considered, the top 20 predictors did not include any of these metrics.

Random-effects meta-analyses were performed for studies that provided combinable effect sizes. All studies were cross-sectional and assessed associations between sleep metrics and anxiety symptoms.

No association between SOL and anxiety symptoms was observed (k=9, N=3643; Fisher z=0.04, 95% CI –0.07 to 0.15; P=.08; Figure 2 [36,39,41,42,48,50-52] and Multimedia Appendix 6), with evidence of moderate to substantial heterogeneity between studies (I²=49.77%). The prediction interval ranged from –0.19 to 0.27. Around half of the included studies were rated as fair quality (k=5, 55.6%). The results remained unchanged in sensitivity analyses based on anxiety type (trait and state) and young adults (Multimedia Appendix 6). One study was not eligible for meta-analysis but found that higher SOL was associated with more anxiety symptoms [45].

No association between TST and anxiety symptoms was observed (k=12, N=14,721; Fisher z=0.009, 95% CI –0.01 to 0.03; P=.28; Figure 3 [36,39,41,42,45,48,50,52,55,56,58] and Multimedia Appendix 6), with very low heterogeneity between studies (I²=3.89%). The prediction interval ranged from –0.02 to 0.03. The majority of the studies were rated as fair quality (k=9, 75%). The results remained unchanged in sensitivity analyses on anxiety type and young adults (Multimedia Appendix 6). Five studies were not eligible for quantitative synthesis. Findings from 4 studies aligned with those from the meta-analysis [38,46,53,66]. One study found a positive association between poor sleep—as measured by a composite including TST and other sleep metrics—and anxiety symptoms [43]. One study reported no association [49].

No association between SE and anxiety symptoms was observed (k=9, N=3710; Fisher z=–0.07, 95% CI –0.14 to 0.002; P=.06; Figure 4 [36,39,41,42,50-52,56] and Multimedia Appendix 6), with low heterogeneity between studies (I²=20.97%). The prediction interval ranged from –0.19 to 0.05. The majority of studies were rated as fair quality (k=7, 77.8%). These results remained unchanged in sensitivity analyses on anxiety type and young adults (Multimedia Appendix 6). Of the 5 studies ineligible for quantitative synthesis, 3 found no association between SE and anxiety symptoms [45,49,57], 1 found a positive association [66], while 2 studies reported negative associations [43,77].

No association between WASO and anxiety symptoms was observed (k=6, N=3291; Fisher z=0.13, 95% CI –0.04 to 0.30; P=.11; Figure 5 [36,39,48,50-52] and Multimedia Appendix 6), with low to moderate heterogeneity between studies (I²=40.62%). The prediction interval ranged from –0.15 to 0.41. Most studies were rated as fair quality (k=4, 66.7%). These results remained unchanged in sensitivity analyses on anxiety type and young adults (Multimedia Appendix 6). One study that was not eligible for quantitative synthesis also found no association [49].

The Hartung-Knapp-Sidik-Jonkman–adjusted CIs reflect uncertainty around the average associations across studies, whereas the prediction intervals estimate the range of effects that may be observed in future studies conducted in different populations or settings. Across the sleep metrics examined, prediction intervals were generally similar to, but slightly wider than, the CIs, suggesting that effects observed in future studies are likely to be close to the average pooled estimates.

This review aimed to advance scientific understanding of the association between digital biomarkers of health collected from wrist-worn devices and anxiety. Specific interest centered on determining which digital biomarkers of health have been the focus of existing research, and which demonstrate the strongest evidence of an association with anxiety. A total of 36 studies were included in this systematic review. Meta-analyses were conducted for 4 sleep metrics (SE, WASO, TST, and SOL), and none of these metrics were associated with anxiety. It was not possible to conduct meta-analyses for physical activity and heart rate metrics; however, qualitative synthesis suggests that lower physical activity levels and elevated heart rate were associated with higher levels of anxiety.

Our meta-analytic findings showed that SE, WASO, TST, and SOL were not statistically associated with anxiety. However, the association for SE was close to the significance threshold, suggesting a possible relationship. Heterogeneity was low for TST and SE, indicating that studies included in these meta-analyses found relatively consistent findings. In contrast, heterogeneity was moderate for WASO and SOL, suggesting that results varied across studies, potentially due to differences in populations, devices used to measure sleep metrics, or anxiety measurement. Subgroup analyses were not conducted as this was not prespecified in the protocol, and the number of studies contributing to each meta-analysis was small [78]. Finally, these findings should also be interpreted in the context of study quality. Most studies were rated as fair quality, reflecting some methodological limitations such as the lack of sample size calculation and no adjustment for key potential confounding variables.

The Pittsburgh Sleep Quality Index, commonly used to assess subjective sleep, comprises 7 components (sleep quality, sleep latency, sleep duration, SE, sleep disturbance, use of sleep medication, and daytime dysfunction) [79]. Existing research using the Pittsburgh Sleep Quality Index generally reports that shorter sleep duration, poor sleep quality, lower SE, and longer SOL are associated with anxiety disorders in both young and older adults [80-83]. In addition, worse subjective sleep quality has also been associated with more nonclinical levels of anxiety symptoms [84,85].

Research examining associations between anxiety and sleep metrics derived from polysomnography, the gold standard measurement of objective sleep, has been limited. One review identified 6 studies investigating polysomnographic sleep metrics, and found decreased SE and TST, and increased WASO in individuals with GAD compared with healthy controls [86]. Another review examined sleep metrics derived from electroencephalography and found that individuals with GAD had decreased TST, and increased WASO and SOL compared with healthy controls; however, no differences in SE were observed [16].

The lack of statistically significant associations observed in our meta-analyses contrasts with some findings from existing literature and may reflect differences in devices and methodological approaches. For instance, only 2 studies included in our meta-analysis asked participants to press an event marker to indicate when they began to fall asleep [36,87]. Without these event markers, estimating SOL may be less accurate [88], as different sleep behaviors can influence wrist movements. The majority of the studies (12/13, 92.3%) in our meta-analyses assessed anxiety symptoms, whereas the studies using polysomnography/electroencephalography focused on individuals with clinical anxiety. The lack of associations in our meta-analyses could also be attributed to differences in sample characteristics. Associations between sleep metrics and anxiety may emerge in clinical anxiety, but not in subthreshold anxiety symptoms.

Our review suggests that higher levels of physical activity are associated with lower anxiety symptoms. These findings appear to be largely consistent across different aspects of physical activity, including step count, total physical activity, and physical activity intensities. These findings align with existing reviews showing the protective effects of physical activity on reducing anxiety symptoms [89,90].

Our review found that sedentary behavior was associated with greater anxiety symptoms. The observed positive association between sedentary behavior and anxiety symptoms aligns with existing findings [91,92]. A meta-analysis primarily based on studies that used self-report questionnaires found a small positive association between sedentary behavior and anxiety [92]. Similar to our review, their study populations were largely composed of healthy individuals rather than those with anxiety disorders. It may be that stronger associations emerge with more severe anxiety. For example, 1 study included in our review observed a significant association only when assessing clinical vs nonclinical anxiety symptoms [46]. No association was observed when analyzing anxiety as a continuous variable (with a mean anxiety score below the clinical threshold) [46]. This potential threshold effect is further supported by another study included in our review, which found that individuals with more severe anxiety symptoms engaged in more sedentary behavior than those with no significant anxiety symptoms [40].

Regarding physical intensities, the majority of studies examining light physical activity (k=3, 75%) and MVPA (k=3, 75%) found that spending more time in these intensities was associated with lower anxiety symptoms. Similarly, a meta-analysis of physical activity interventions found that these interventions reduced anxiety symptoms [93]. Although the number of studies within each subgroup is limited, subgroup analyses of exercise intensity found that both moderate- and high-intensity exercise reduced anxiety symptoms [93]. Taken together, our review’s findings on physical activity intensities align with existing research showing that higher-intensity physical activity is effective in reducing anxiety symptoms.

Our review found inconclusive evidence on the association between HRV and anxiety, as only 4 studies included in this review examined HRV. An existing meta-analysis on HRV found that individuals with GAD had reduced high-frequency and time-domain HRV [94], which means they may have less flexibility in adapting to stress. This partially aligns with the findings from 2 studies included in this review: 1 found that higher HRV (domain not specified) was associated with anxiety symptoms [66], while another found that higher mean time-domain HRV was associated with lower anxiety symptoms [61]. This study also found that greater high-frequency HRV variability was associated with more anxiety symptoms [61]. The discrepancy in findings may be due to the focus of the existing meta-analysis on clinical anxiety, whereas the 2 studies in our review examined anxiety symptoms.

Elevated resting or average heart rate reflects chronic SNS activation (SNS regulates the body’s fight-or-flight response) [95]. Compared to HRV, far fewer studies have examined average heart rate or resting heart rate in relation to anxiety. Existing evidence suggests that a higher resting heart rate is linked to greater anxiety symptoms and anxiety disorders [96,97]. Consistent with this literature, our review found that higher average heart rate and resting heart rate were associated with more anxiety symptoms.

Machine learning approaches emphasize building models that generalize to new data. These approaches offer advantages over inferential statistical methods [98]. For example, machine learning approaches can capture both linear and nonlinear relationships between variables without strong assumptions about variable relationships. However, machine learning models can be challenging to interpret due to their “black box” nature, where decision-making processes are not transparent [99].

In this review, machine learning models predicting anxiety had accuracies ranging from 56.3% to 90.9%. The accuracy reported here is somewhat comparable to the 2 reviews on machine learning models using any wearable devices to predict depression (70%-89%) [17] and anxiety (82%) [18], which used a range of predictors (eg, heart rate, activity, sleep, audio, and skin temperature). The majority of studies included in this review did not include demographic predictors, which might further enhance model accuracy. Indeed, some studies in the depression and anxiety machine learning reviews used demographic data as predictors, further supporting the potential use of these predictors to improve the prediction of anxiety. The model with the highest performance (accuracy of 90.9%) combined digital biomarkers of health from more than 1 type of digital device (wrist-worn wearable and smart shirts) [72]. This aligns with existing research showing that models integrating digital biomarkers of health from multiple devices achieved better predictive performance than those relying on a single device [20]. For studies using only digital biomarkers of health from wrist-worn wearables in this review, the models with the highest accuracy included anxiety questionnaires as a predictor to predict clinically diagnosed anxiety. This likely inflates model performance due to circularity. When excluding models that incorporated anxiety questionnaires as inputs, the studies using only digital biomarkers of health from wrist-worn wearables reported accuracies ranging from 56.3% to 69.4%.

Digital biomarkers from wrist-worn wearables offer several benefits over traditional assessments of anxiety, as they allow for continuous data collection over long periods without the need for repeated administration. In addition, these largely affordable and minimally intrusive devices can be used in natural settings, providing valuable insights into sleep and activity habits. While both research-grade and consumer-grade devices are less precise than polysomnography in measuring sleep [100-102], this may be due to their indirect measurement (eg, based on movement). Despite this limitation, they are generally considered valid tools for sleep assessment [100,103]. Moreover, polysomnography is neither cost-effective nor practical for large-scale real-world studies. Given that the majority of the studies in our review used research-grade devices, a key remaining question is whether these findings generalize to consumer-grade wrist-worn devices, which are more widely accessible to the general population.

There are also benefits to using self-report data. Self-report questionnaires can capture subjective experiences, such as perceived levels of exercise or sleep quality, which wearables cannot measure. Further, they have the potential to record underlying reasons for poor sleep, for example, bad dreams or pain. Therefore, while wrist-worn wearables offer advantages, incorporating self-report measures can provide a more comprehensive understanding of both objective and subjective experience.

To the best of our knowledge, this is the first review to examine the association between digital biomarkers of health from wrist-worn wearables and anxiety. We used double screening for all studies to ensure a rigorous selection process and to minimize bias. We also included gray literature to reduce potential publication bias and provide a more comprehensive review of existing studies.

This review also has several limitations. First, most studies focused on young adults. Part of the impetus for this review was to explore whether metrics from wrist-worn wearables could eventually be used to screen for dementia risk, as indicated by changes in anxiety symptoms. This is because anxiety is a potentially modifiable factor associated with dementia risk. However, the number of studies conducted in older adults is scarce. Future studies involving older adults would enrich our understanding of the association between digital biomarkers of health and anxiety in the context of dementia. Second, despite previous literature showing associations between physical activity and anxiety [104,105], few studies in our review focused on activity-related metrics. This precluded a meta-analytic synthesis of these findings. Wrist-worn devices generally provide more objective measurements of activity compared to self-report, which can be influenced by different biases [106], highlighting the need for further research. Third, although our search strategy and eligibility criteria allowed inclusion of any metrics measured by wrist-worn wearables, the majority of studies focused on sleep metrics. The search strategy was not peer-reviewed, and although it is possible that consulting with an information scientist might have led to the inclusion of additional relevant studies, this gap highlights an opportunity for future research to explore a wider range of metrics that could be associated with anxiety. Fourth, while machine learning has potential in predicting anxiety, the quality of these studies needs improvement. For example, future studies need to carefully consider the predictors used. None of the included machine learning studies considered demographics as potential predictors, despite existing research indicating sex differences in anxiety [107]. In addition, future studies should increase the interpretability of machine learning models by routinely reporting Shapley additive explanations values or feature importance scores. Additionally, most of the studies focused on a single metric (eg, sleep), despite the potential for the models to include multiple predictors. More research using machine learning approaches is needed, especially to compare whether this approach is superior to traditional approaches that use inferential statistics in predicting anxiety.

Wrist-worn wearable devices offer a scalable and noninvasive method to passively and continuously monitor anxiety symptoms. Metrics collected from these devices have the potential to serve as a screening or detection tool for anxiety, allowing individuals at risk of developing clinical anxiety to be identified before consulting a clinician. If these metrics and models have high predictive accuracy, they could support clinicians in identifying at-risk individuals and provide timely assessments and interventions. Before such an approach can be implemented in real-world clinical practice, more research is essential. In order to develop a clinically meaningful prediction model, interdisciplinary collaboration is crucial (eg, clinicians or machine learning experts). This will ensure the models are accurate, interpretable, clinically meaningful, and integrated into the health care system rather than remaining a research-focused tool.

This review is the first to comprehensively synthesize evidence on the association between digital biomarkers of health derived from wrist-worn wearable devices and anxiety, integrating findings from both inferential statistical and machine learning studies. By systematically synthesizing the literature, this review identified gaps in evidence for future research.

Our meta-analyses found no associations between SOL, TST, SE, and WASO and anxiety. Although the number of studies was limited, there was evidence that lower physical activity levels and elevated heart rate were associated with more anxiety symptoms. Existing studies on light physical activity, MVPA, and HRV present inconsistent or inconclusive findings, highlighting the need for more research. Further, a comprehensive review of the literature yielded few machine learning studies in this area, and their quality was variable. While machine learning is increasingly used to analyze wearable data, its application for predicting anxiety remains in the early stages.

Machine learning and digital biomarkers could become powerful tools for early detection of anxiety, symptom monitoring and management, and prevention, particularly when used as part of broader assessment frameworks. Continuous data collection could support long-term monitoring of behavioral and physiological patterns relevant to anxiety, and may complement self-report and clinical assessments rather than being used as stand-alone screening tools. Given that anxiety is a potentially modifiable psychological factor associated with dementia risk, wearable-derived digital biomarkers may also offer future opportunities to support early identification and monitoring of dementia risk. However, further high-quality research is required before such approaches can be meaningfully translated into clinical or public health practice.