Wearable artificial intelligence for anxiety and depression: A scoping review

.


Table of Contents
No, I do not wish to publish my submitted manuscript as a preprint.2) If accepted for publication in a JMIR journal, would you like the PDF to be visible to the public?
Yes, please make my accepted manuscript PDF available to anyone at any time (Recommended).
Yes, but please make my accepted manuscript PDF available only to logged-in users; I understand that the title and abstract will remain v Yes, but only make the title and abstract visible (see Important note, above).I understand that if I later pay to participate in <a href="http

INTRODUCTION Background
Anxiety and depression are amongst the "common mental illnesses" with high global prevalence.As of 2020, it has been reported that 19 percent of people worldwide suffered with depression or anxiety that prevents them to do their regular daily activities as they usually would for two weeks or longer 1 .In addition to having a significant economic impact on society 2 , anxiety and depression affect people in terms of lost years because of illness.The statistics are mind blowing, depression is the world's leading cause of disability within the youth population 3-5 .At 18 years of age, a previous study observed that depressed adults had 28 more years of quality-adjusted life expectancy (QALE) than non-depressed adults, resulting in a 28.9-year QALE loss due to depression in the United States 6 .Depression is also a significant risk factor when it comes to suicide 7 .The abovementioned statistics combined with the fact we only have around 9 psychiatrists per 100,000 people in developed countries 8 and 0.1 for every 1,000,000 in low-income countries 9 the situation is challenging to say the least.Current approaches for the assessment of anxiety and depression disorders are primarily based on clinical observations of patients' mental states, clinical history, and self-report questionnaires, such as the General Anxiety Disorder-7 (GAD-7) for anxiety and Patient Health Questionnaire-9 (PHQ-9) for depression.These methods are subjective, timeconsuming, and challenging to repeat.As a result, contemporary psychiatric assessments can be inaccurate and ineffective at assessing anxiety and depression symptoms in a reliable and personalized manner.Therefore, there is a significant need to develop automatic techniques to address the limitations of the current psychiatric approaches for assessing anxiety and depression disorders and to overcome the shortages and uneven distribution of mental health professionals.
Recently, there have been rapid ongoing developments of artificial intelligence (AI) technology and wearables technology for healthcare and clinical use, offering numerous advantages towards individualizing diagnoses and treatment management of psychiatric disorders, including anxiety and depression 10-12 .Wearable technology includes electronic devices which users can wear near-body (e.g., smart watch, smart glasses, smart bracelet), onbody (e.g., electrocardiogram electrodes), in-body (e.g., implantable smart patch), and electronic textiles (e.g., smart clothes).Wearable devices are designed to provide a constant stream of health care data for disease diagnosis and treatment.This is achieved by continuously recording physiological parameters such as temperature, blood pressure, blood oxygen, respiratory rate, physical movement, and the electrical activity of the heart, brain, and skin.Symptoms of anxiety and depression can be assessed by many parameters collected in real-time by wearable devices for the diagnosis and monitoring of patients with anxiety and depression.However, the dramatically accelerating pace in the development and adoption of wearables coupled with a shortage of skilled caregivers has led to an evolving need for automatic, efficient, and real-time approaches to analyze the large volumes of data collected by wearable sensors.This has motivated the integration of AI methods into wearable devices, introducing the "Wearable AI" technology.Wearable AI refers to intelligent electronic devices which are designed to be worn on the user's body with intelligent operations.Wearable devices typically deal with monitoring and analyzing patients' health data.However, when paired with AI, wearable AI introduces fundamental developments in the diagnosis and treatment of anxiety and depression.It has the potential to provide an early and accurate diagnosis of anxiety and depression, facilitate more individualized treatment for anxiety and depression patients, and assist in developing preventative measures for groups at risk of anxiety and depression.

Research Problem and Aim
An extensive number of studies have been published on wearable devices combined with AI for anxiety and depression.Several reviews were conducted to summarize previous studies; however, they had the following limitations.Firstly, they focused on wearable devices rather than wearable devices paired with AI 10-15 .Secondly, they did not describe in detail the features of the used wearable devices and AI models 10-15 .Thirdly, they only targeted certain age groups such as children and adolescents 10,12 .Fourthly, they focused on wearable devices for either anxiety 11,14 or depression 12,13,15 rather than both anxiety and depression.Fifthly, they did not search relevant databases such as Medline 14 , PsychInfo 10,13,15 , IEEE Xplore 10-14 , ACM Digital Library 10-15 .Lastly, they focused on wearables devices used for only diagnosing purposes using only ECG data 11 or EEG data 15 .Therefore, the need for a review that focuses on AI-paired wearable devices for anxiety and depression has never been higher.The review should be the same high-quality of a previous review conducted about AI-paired wearable devices for diabetes 16 .The current review aimed at exploring the features of wearable AI used for anxiety and depression, both to help customers make educated selections and to help the research community advance in this field by identifying gaps and looking into future prospects.

METHODS
To achieve the objective of the study, we carried out a scoping review consistent with Preferred Reporting Items for Systematic Reviews and Meta-Analyses-Extension for Scoping Reviews (PRISMA-ScR) 17 .PRISMA-ScR Checklist for this review is presented in Multimedia Appendix 1.The methods used in this review are detailed in the following subsections.

Search strategy
To find relevant studies, we searched 8 electronic databases on May 30, 2022: MEDLINE (via Ovid), PsycInfo (via Ovid), EMBASE (via Ovid), CINAHL (via EBSCO), IEEE Xplore, ACM Digital Library, Scopus, and Google Scholar.We set an automatic search biweekly for 24 weeks (ending on September 30, 2022).Given that Google Scholar retrieved a massive number of hits and order them based on their relevancy, only the first 100 hits (i.e.,10 pages) were checked in this review.To identify additional studies, we checked the reference lists of included studies (i.e., backward reference list checking) and screened studies that cited the included studies (i.e., forward reference list checking).
To develop the search query, three experts in digital mental health were consulted and previous reviews of relevance to the review were checked.The search query was composed of 3 groups of terms: terms related to AI (e.g., artificial intelligence, machine learning, and deep learning), terms related to wearable devices (e.g., wearable OR smart watch OR smartwatch), and terms related to anxiety and depression (e.g., anxiety OR anxious OR depression).Multimedia Appendix 2 presents the detailed search query used for searching each database.

Study Eligibility Criteria
This review included studies that focused on developing AI algorithms for anxiety and depression using data collected by wearable devices.Specifically, we focused on all AI algorithms used for any purpose related to anxiety and depression (e.g., diagnosis, monitoring, screening, therapy, predication, and prevention).The wearable devices that were used for collecting data had to be non-invasive on-body wearables such as smartwatches, smart glasses, smart clothing, smart bracelets, and smart tattoos.On the other hand, we excluded studies that used data collected by the following devices: non-wearable devices, hand-held devices (e.g., mobile phones), near-body wearable devices, in-body wearable devices (e.g., implants), wearable devices connected with non-wearable devices using wires, and wearable devices that need an expert to apply on users (e.g., wearable devices composed of many electrodes that need to be placed in very specific points of the body).Studies that used data collected via any methods (e.g., non-wearable devices, questionnaires, and interviews) in addition to wearable devices were considered in this review.We excluded studies that showed only a theoretical framework of AI-based wearable devices for anxiety and depression.We included journal articles, conference papers, and dissertations that were published in the English language since 2015.We excluded reviews, preprints, conference abstracts, posters, protocols, editorials, and commentaries.No restrictions were applied regarding the measured outcomes, setting, and country of publications.

Study Selection
We followed three steps in the study selection process.In the first step, we used EndNote X9 to remove duplicates from all retrieved studies.In the second step, we checked the titles and abstracts of the remaining publications.Lastly, we screened the entire texts of the studies included in the previous step.Two reviewers independently performed the study selection process.Disagreements between them in the second and third steps were resolved by discussion.Cohen's kappa was calculated to measure the inter-rater agreement 18 , and it was 0.85 for "title and abstract" screening and 0.92 for full-text reading.

Data Extraction
Two reviewers utilized Microsoft Excel to independently extract data about study meta-data, wearable devices, and AI techniques.Any disagreements between the reviewers were resolved through discussion.The data extraction form used in this review was piloted using 5 studies, and it is shown in Multimedia Appendix 3.

Data Synthesis
Data that was extracted from the included studies were synthesized using the narrative approach, where data was summarized and described using texts, tables, and figures.To be more specific, we started by describing the meta-data of the included studies (e.g., year of publication and country of publication).Then, we presented the features of wearables devices used in the included studies (e.g., their status, type, placement, and operating system).Lastly, we summarized the characteristics of AI techniques used (e.g., AI algorithms used, their aim, dataset size, and data input type).We used Microsoft Excel to manage data synthesis.

Search Results
As depicted in Figure 1, searching all pre-identified databases retrieved 1203 records.Of these, 340 duplicates were detected and removed using reference management software (EndNote X9).Screening titles and abstracts of the remaining 863 citations resulted in excluding 506 records.We could find the full text of 7 records of the remaining 357 records.
Reading the full text of the remaining 354 records led to excluding 298 records for several reasons shown in Figure 1.We identified 13 additional records relevant to this review by backward and forward reference list checking.In total, 69 records were included in the current review 19-87 .
3 number of studies does not add up as several studies used more than one wearable device, and many wearable devices are compatible with more than operating system (OS). 4number of studies does not add up as several studies used more than one wearable device, and many wearable devices used more than one gateway. 5number of studies does not add up as several studies used more than one wearable device, and many wearable devices used more than one host. 6number of studies does not add up as several studies used more than one wearable device, and many wearable devices used more than one of mode of data transfer.
2 number of studies does not add up as several studies used more than one wearable device and most wearable devices have more than one sensor. 3number of studies does not add up as several studies used more than one wearable device and many wearable devices used more than sensing approach. 4number of studies does not add up as several studies used more than one wearable device and many wearable devices used more than sensing type.
2 number of studies does not add up as many studies used more than one AI algorithm.
3 number of studies does not add up as many studies used more than one tool to assess the ground truth. 4number of studies does not add up as many studies used more than one validation approach. 5number of studies does not add up as most studies used more than one performance measures.
2 number of studies does not add up as several studies used various numbers of features.

Principle Findings
This scoping review aimed at exploring features of AI and wearable devices used for anxiety and depression.In this review, about two thirds of the studies used wearable AI for depression while the remaining studies used it for anxiety.This may be attributed to the capabilities of wearables to collect biosingals related to symptoms of depression and anxiety.More specifically, it is well known that depression is associated with a decrease in activity and changes in sleep behaviours 13,89,90 , which can be objectively measured by wearable devices.Further, analysis of depression symptoms does not rely upon highly accurate data; that is, general trends are sufficient to provide indications.In contrast, anxiety is usually associated with heart rate variability 91 .Although wearable devices can have an acceptable heart rate accuracy 92 , the quality differs among devices 93 .Beyond, monitoring the heart rate without context information might be misleading since multiple factors impact the heart rate, thus, detecting anxiety based on only objective biosingals is questionable.Combination with additional data sources is crucial.So far, only a few studies in this review are based upon a combination of data from different sources (i.e., wearable devices, non-wearable devices, and self-administered questionnaires).
In this review, the most frequent application of wearable AI is diagnosing or screening anxiety and depression.A similar result was reported by 2 previous reviews, which showed that most studies focused on using wearables for diagnostic purposes 10,13 .Although wearable AI can be used for interventional and treatment purposes (e.g., personalized mindfulness, meditation, and biofeedback therapy 14 ), none of the systems in included studies was used for such purposes.This may be attributed to the lack of evidence on the effectiveness of wearable AI for improving anxiety and depression.Smart bands worn on the wrist were most often applied in the studies.This has already been indicated by previous reviews as well 10,13,14 .This can may be attributed to the fact that wristworn wearable devices are less distractive and obtrusive, easy to use, and more stylish and familiar to most people.According to Hunkin et al. 94 , such features are crucial for users' acceptance and use of wearable devices.
The most commonly used data for model development were physical activity data, sleep data, and heart rate data.This is expected given that depression and anxiety are associated with physical activity 13,89,90 , sleep pattens 13,95,96 , and heart rate 91 , in addition, as the current review showed, these are the most common biosignals measured by commercial wearable devices.Surprisingly, more than half of the papers considered only data from wearables in their AI algorithms.However, wearables cannot detect all symptoms of relevance for anxiety and depression for 2 reasons.Firstly, wearable devices cannot detect several physiological data such as weight loss or gain and changes in appetite 13 .Secondly, wearable devices cannot evaluate subjective symptoms such as social interaction, medical history, and lifestyle changes 13 .We might question whether research starts to place overreliance upon the diagnostic and predictive power of data from wearable devices only.About one-fourth of studies relied upon a dataset called Depresjon 35 to develop their models.Depresjon is a freely available dataset that contains data related to the motor activity measured using an actigraph watch worn at the wrist (Actiwatch AW4, Cambridge Neurotechnology Ltd) 35 .The dataset also contains data related to depression levels assessed using the MADRS 35 .This explains why the most common wearable device used in the included studies was Actiwatch AW4 and why MADRS was the most frequently used tool to assess the ground truth.
Regarding the target population, we have to recognize that the majority of studies addressed individuals between the ages of 18 and 65.Global statistics show that depression and anxiety occur all over the age ranges starting at 15 with almost the same percentage.Only for adults at an age of 65 and older, there is a decrease in the percentage 1 .This might explain why the studies mainly targeted the age group 18 to 65.Another explanation might be that wearables are more popular for adults in that age range.This review showed that K-fold cross-validation was the most frequently used validation method.This can be attributed to several reasons.Firstly, in comparison with hold-out crossvalidation, K-fold cross-validation is prone to less variation as each observation is used for both training and testing.Secondly, the training set in K-fold cross-validation is larger than the training set in hold-out cross-validation, thereby, K-fold cross-validation has reduced bias and reduced over-estimation of test-error.Lastly, K-fold cross-validation is less expensive computationally than LOOCV as the algorithm needs to rerun only k times (usually ≤10).

Research and Practical Implications
The performance of wearable AI in diagnosing, monitoring, and predicting anxiety and depression was not assessed in this review.Systematic reviews and meta-analyses are needed to examine its performance.Future studies should also compare the performance of different wearable devices (e.g., Fitbit vs. Empatica), worn at different placements (e.g., wrist, chest, waist), and using different data types (e.g., wearable based-data vs. wearable based-data and self-reported data).Conducting systematic reviews of such studies can help researchers, developers, and wearable device companies to identify the most significant features and powerful AI algorithms in diagnosing, monitoring, and predicting anxiety and depression.AI research highly depends on available datasets.However, when only one dataset is exploited by researchers, no conclusions regarding the generalizability of study results can be drawn.Therefore, we recommend researchers (1)  Accordingly, the performance of AI-based wearable devices will be underestimated.
Although the current studies showed that wearable AI can be used for monitoring symptoms or levels of anxiety and depression, continuous tracking of physiological biomarkers could trigger emotional instability and ruminative thinking 97 .Although the wearable AI can approximate mental states (e.g., feeling nervous, anxious, or on edge) through heart rate and other variables, it could provide many false positives, thereby, exacerbating or increasing the anxiety or depression of an individual.The above-mentioned downsides of wearable AI should be considered and mitigated before developing AI-based wearables.More research studies are needed on the use of wearable devices and their impact on individual emotional and behavioural responses to a wearable device's automated feedback.
Wearable AI can help individuals conduct mental health and well-being pre-screening assessments without an initial hospital or clinical encounter.The individual could be notified through the wearable device, smartphone, or desktop application about their mental health status which would encourage them to visit a mental health and well-being professional.Such pre-screening feedback from wearables may help reduce mental health stigma and allow a higher number of individuals to seek help from a mental health professional.
The quality of the data, whether it is obtained from open sources or generated from wearable devices, should be emphasized.To do so, there is a need to be more practical standards for wearable device development that ensures accurate measurement of different signals generated from wearable devices to improve algorithmic performance.

Limitations
This review excluded many studies that focused on non-wearable devices, hand-held devices (e.g., mobile phones), near-body wearable devices, in-body wearable devices (e.g., implants), wearable devices connected with non-wearable devices using wires, and wearable devices that need an expert to apply on users.For this reason, our findings may not be generalizable to contexts where such excluded devices are applied.Owing to practical constraints, we included only studies published in the English language.We also restricted our search to studies published from 2015 onwards given that this is a fast-growing field, thereby, studies published before 2015 can be deemed outdated.Consequently, it is likely that we missed some studies published in other languages and/or published before 2015.Another limitation of this review is that we cannot comment on the performance of wearable AI in diagnosing, monitoring, and predicting anxiety and depression and the importance of features/variables as this is out of the scope of the current review and needs systematic reviews, where the quality of the evidence and risk of bias are assessed.

CONCLUSION
Wearable AI can offer great promise in providing mental health services related to anxiety and depression.Wearable AI can be used by individuals as a pre-screening assessment of anxiety and depression.Further reviews are needed to statistically synthesize studies' results related to the performance and effectiveness of wearable AI.More studies are needed on the use of wearable devices and their impact on individual emotional and behavioural responses to a wearable device's automated feedback.Given its potential, tech companies should invest more in wearable AI for treatment purposes for anxiety and depression.Downsides of wearable AI (e.g., false positive alerts and triggering emotional instability and ruminative thinking) should be considered and mitigated before developing it.

Figure 1 :
Figure 1: Flow chart of the study selection process

Figures
Figures

Table 2 )
. In 13 of the 21 studies (61.9%), the gateway was PCs, smartphones, and tablets.The included studies used 4 types of host devices (i.e., end gate devices that stores data collected by the wearable devices).More than one host device was used in 14 studies (20.3%).The most common host devices in the included studies were computers (46/69, 66.7%) and database servers (30/69, 43.5%).Data is transferred from the wearable device to the host device through 6 different modes.In about 46.4% (32/69) of the studies, more than one mode of data transfer was used.The most common mode was Bluetooth (41/69, 59.4%) followed by docking stations (27/69, 39.1%) and Internet (24/69, 34.8%).

Table 2 :
Features of wearable devices

Table 4 :
Features of AI algorithms

Table 5 :
).The included studies used datasets from either closed sources (i.e., collected by authors of the study or obtained from previous studies) (50/69, 72.5%) or open sources (i.e., public databases) (19/69, 27.5%).Depresjon was the most common dataset obtained from open sources and used in the included studies(16/19, 84.2%).In 59.4% (41/69) of the studies, AI algorithms were developed using data collected by only wearable devices.Around 17.4% (12/69) of the studies developed AI algorithms using data collected by a combination of wearable devices and self-administered questionnaires (i.e., self-reported data).About 13% (9/69) of the studies developed AI algorithms using data collected by a combination of wearable devices and non-wearable devices (e.g., smartphones).Around 10.1% (7/69) of the studies developed AI algorithms using data collected by a combination of wearable devices, non-wearable devices, and selfadministered questionnaires.The included studies used more than 50 categories of data to develop the model.While 43.5% (30/69) of the studies used only one category of the data to develop their models, the rest of the studies (39/69, 56.5%) used more than one category of the data.The most common data used to develop the models were physical activity data (e.g., step counts, calories, metabolic rate) (53/69, 76.8%), sleep data (e.g., duration, patterns) Features of data used for AI development (27/69, 39.1%), heart rate data (e.g., heart rate, heart rate variability, interbeat interval) (26/69, 37.7%), mental health measures (e.g., depression level, anxiety level, stress level, mood status) (14/69, 20.3%), location data (e.g., latitude, longitude, % of time at home, stationary time) (10/69, %14.5), smartphone usage data (e.g., display on/off, charging activity, number of apps used) (10/69, %14.5), and social interaction (e.g., call and publish their datasets in open databases after ensuring participants' privacy and confidentiality and (2) exploit different datasets available in open databases.The current review found a lack of AI-based wearable devices used for treatment purposes although wearable AI can be used for providing many interventions for anxiety and depression such as personalized mindfulness, meditation, and biofeedback therapy.Tech companies should invest more in wearable AI for treatment purposes for anxiety and depression.Researchers should also assess the effectiveness of such technologies in improving anxiety and depression.The ground truth of mental states (anxiety or depression) in included studies was identified based on 27 different tools.Although most of these tools have been validated extensively, they usually do not include physiological biomarkers (e.g., physical activities, heart rate, EDA, respiratory rate, EEG).This brings into question the validity and reliability of drawing conclusions about mental states (anxiety or depression) based on physiological biomarkers when the grand truth of mental states is assessed using subjective questionnaires.