Sleep is a fundamental biological process essential for maintaining overall health and well-being. It is a dynamic state in which the brain processes daily experiences, promotes synaptic plasticity, and supports physical functions. During sleep, the brain and body engage in recovery, repair, and preparation for the next day [1]. Sufficient sleep is crucial for mood stability, cognitive function, and overall health. Both sleep quantity and quality are vital for optimal functioning of the body and mind [2]. The National Sleep Foundation defines optimal sleep quantity for adults as 7-9 hours per night [3], while sleep quality is characterized by factors such as minimal interruptions, appropriate sleep onset latency (typically under 30 minutes), and a significant proportion of restorative sleep stages (eg, deep sleep, rapid eye movement sleep). According to the Philips Global Sleep Survey [4], 62% of people worldwide report not getting the quality of sleep they desire, and 44% have experienced worsening sleep over the past 5 years, a problem that may be attributed to various sleep disorders. The International Classification of Sleep Disorders categorizes sleep disorders into insomnia, sleep-disordered breathing, central hypersomnolence disorders, circadian rhythm sleep-wake disorders, parasomnias, and sleep-related movement disorders [5].

Existing research has shown that sleep disorders significantly impact both physical and mental health. They can manifest as insufficient sleep, excessive sleep, or abnormal movements during sleep. Several studies have found that sleep disorders are associated with an increased risk of cardiovascular disease, diabetes, and cancer [6-8]. Additionally, they are linked to mental health issues such as depression, anxiety, and suicidal behavior [9-11]. Beyond individual health, sleep disorders also have broader societal consequences, including an increased risk of road accidents [12]. To mitigate the negative health and social impacts of sleep disorders, early detection, monitoring, and treatment are essential.

Various methods and devices have been used to monitor and diagnose sleep disorders, including polysomnography (PSG), home sleep testing (HST), and actigraphy. PSG is the gold standard for diagnosing sleep disorders, as it accurately assesses sleep phases and identifies potential conditions. However, despite its advantages, PSG has some limitations. It is costly and time-consuming, requires individuals to spend the night in a sleep laboratory, and depends on expert monitoring and scoring. By contrast, HST and actigraphy are less costly, allow data collection over multiple days, and can be used in nonlaboratory settings compared with PSG [13,14]. However, HST has limitations, including underestimating results and providing limited evaluations of certain sleep disorders [15,16]. These limitations can be addressed by wearable artificial intelligence (AI) technology.

As wearable AI devices become increasingly popular, they have revolutionized the health care industry by enabling real-time monitoring and diagnostic capabilities [17]. This technology integrates AI into wearable devices (WDs), allowing them to perform tasks such as data processing, inference, and decision-making directly on the device [18]. According to the IEC (International Electrotechnical Commission) Standardization Group 10, wearable smart devices are categorized into 4 groups based on their proximity to, placement on, or implantation within an organism, such as the human body (as cited in [19]). Near-body wearables, such as radar-based monitoring systems, contactless sleep-tracking devices, and mobile sleep apps, operate close to the body but do not require direct skin contact. On-body wearables, including smartwatches, fitness trackers, smart glasses, electrocardiogram electrodes, electromyography sensors, and electrodermal activity monitors, are worn directly on the body and maintain continuous skin contact. In-body wearables, such as implantable smart patches and pacemakers, are implanted into the body. Electronic textiles integrate fabric-based electronics, including smart clothing designed to monitor physiological parameters.

Several studies have been published on the use of WDs combined with AI to detect or monitor sleep disorders. While multiple reviews have summarized previous studies, certain limitations exist. Some reviews focused solely on the features of AI models without discussing their integration with WDs [20-22]. Others examined only a specific type of sleep disorder [23-25]. Additionally, several reviews used search queries that omitted important terms [22,26]. Numerous reviews did not include searches in popular databases such as MEDLINE, PsycINFO, and Embase [20-22]. Some reviews focused on specific types of data, such as clinical data or consumer data from sleep technology devices used outside clinical settings [20,26]. Several reviews were narrative in nature, indicating that they did not follow systematic approaches [22,26]. Therefore, this review aims to provide an overview of AI-powered WDs used for sleep disorders by analyzing key aspects across 3 dimensions. First, it examines the study characteristics, including design, population, and geographical trends, to highlight research patterns. Second, it explores the technological features of WDs, such as sensor types and biosignals collected, emphasizing their role in sleep monitoring. Third, it investigates the AI methodologies employed, their applications, and validation approaches, showcasing advancements in AI-driven sleep disorder detection.

To ensure a thorough and systematic approach, this scoping review adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews). A detailed account of adherence to PRISMA-ScR guidelines is provided in Multimedia Appendix 1, outlining the structured process we followed.

A comprehensive search was conducted across several electronic databases, including MEDLINE, PsycINFO, Embase, IEEE Xplore, ACM Digital Library, Scopus, and Google Scholar. An automatic alert was set to run the search query biweekly. The bibliographic data collection period spanned from December 7, 2023, to March 6, 2024. As a result of the overwhelming volume of results from Google Scholar and its ability to prioritize relevant search results, this review deliberately focused only on the first 100 results. To identify additional relevant sources, we performed backward and forward reference list checking. This process involved analyzing the reference lists of included articles and examining studies that cited them using Google Scholar’s “Cited by” feature.

The search queries combined terms related to sleep disorders (eg, sleep disorder*s, sleep disturbance, and sleep apnea) with terms related to AI (eg, AI, machine learning, and deep learning) and WDs (eg, wearable, smartwatch, and smart band). In collaboration with digital health experts and after reviewing relevant literature, the final search query was meticulously crafted. Boolean operators “OR” and “AND” were used to combine terms within the same category and across different categories, respectively. The language filter was set to English only. Duplicates were identified and removed using EndNote X9 (Clarivate Plc). Full details of the search terms used for each electronic database are provided in Multimedia Appendix 2.

This review encompassed studies that utilized AI algorithms for any purpose related to sleep disorders using data from WDs. Research articles were deemed suitable for inclusion if they primarily focused on individuals diagnosed with or suspected of having any type of sleep disorder, without restrictions based on age, gender, or ethnicity. Studies that focused solely on AI applications for detecting sleep quality or sleep staging—without directly addressing sleep disorders—or those forecasting intervention outcomes for sleep disorders were excluded.

This review included studies that gathered data using noninvasive, on-body WDs. Research papers that exclusively relied on non-WDs, handheld devices (eg, mobile phones), near-body or in-body WDs, WDs physically connected to non-WDs, or wearables requiring expert oversight—such as those necessitating precise electrode placement—were excluded. Studies on animals or patients with other primary health conditions were also eliminated. Additionally, only peer-reviewed journal articles, conference papers, and dissertations were considered, with no restrictions on study setting, study design, reference standard (ie, ground truth), year of publication, or country of study. However, papers not published in English or classified as editorials, preprints, reviews, protocols, posters, conference abstracts, or research highlights were excluded from consideration.

The study selection process in this review comprised 2 key steps. First, all retrieved articles underwent a preliminary screening based on their titles and abstracts by 2 reviewers. This step was essential for determining whether the articles met the inclusion criteria without requiring a full-text review. It aimed to exclude studies that clearly did not meet the criteria, such as those unrelated to WDs or focusing on other aspects of sleep technology.

Articles that passed the initial screening were then subjected to a detailed full-text review. The same 2 reviewers independently conducted this assessment, thoroughly evaluating each study against the inclusion and exclusion criteria to confirm its relevance to the research questions. Studies that lacked sufficient data on AI algorithm performance, used nonwearable technology, or fell outside the scope of peer-reviewed literature were excluded. Any discrepancies between reviewers were resolved through discussion until a consensus was reached. If disagreement persisted, a third reviewer was consulted to make the final decision.

Two reviewers independently extracted data from the included studies using Microsoft Excel. The extracted information included study metadata, WD features, and AI algorithm characteristics. The data extraction form used in this review is provided in Multimedia Appendix 3. Any differences in data interpretation or extraction between reviewers were resolved through discussion until a consensus was reached.

We used a narrative approach to synthesize the extracted data, which were then aggregated using text, tables, and figures. Specifically, we first presented the search results, followed by an overview of the studies’ general characteristics, and finally, a detailed description of the features of WDs and AI technologies. We examined the technical characteristics of WDs, including key measurements, sensing approaches, and sensor properties, as well as general attributes such as device type, placement, and status. The AI aspects were analyzed based on the models used, evaluation criteria, and their applications.

Figure 1 illustrates the study selection process, as per PRISMA-ScR guidelines (Multimedia Appendix 1). The initial database search yielded 689 citations. After identifying and removing 240 duplicates using EndNote X9, 449 unique studies remained. Screening the titles and abstracts led to the exclusion of 397 studies. The full texts of the remaining 52 studies were retrieved and assessed, resulting in the exclusion of 9 studies. The primary reasons for exclusion were the lack of studies focusing on sleep disorders (n=2), the absence of AI algorithms (n=6), and inappropriate publication type (n=1). Additionally, 3 relevant studies were identified through reference list screening. Ultimately, 46 studies were included in this review.

As shown in Table 1, the number of studies fluctuated over time, with the highest counts recorded in 2023 and 2020 (11/46, 24%). The included studies were conducted across 17 different countries, with the United States contributing the most (10/46, 22%). The majority of the research was published as journal articles (36/46, 78%). The average number of participants per study was 218.4 (SD 597.2), ranging from 4 to 3414. Among the 29 studies that reported participant ages, the age range was 12-68 years, with an average of 45.8 (SD 12.4) years. The proportion of female participants across 30 studies averaged 39.2%, ranging from 12% to 65%. The majority of studies (42/46, 91%) focused on sleep apnea. Multimedia Appendix 4 provides the characteristics of each included study.

Commercial WDs constituted the majority of the included studies (30/46, 65%; Table 2). The most frequently mentioned WDs were Actiwatch, Belun Ring, and Fitbit (3/46, 7%), with smart bands being the most common type (12/46, 26%). WDs were placed on various body parts, with the wrist (19/46, 41%), chest (15/46, 33%), and abdomen (8/46, 17%) being the most common locations. Most of these devices collected activity and sleep measures (10/46, 22%), along with other biosignals. As illustrated in Figure 2A, the most common sensors identified in these WDs were accelerometers (34/46, 74%) and photoplethysmography sensors (14/46, 30%). Figure 2B highlights a clear trend of accelerometer sensor adoption over the years, often in combination with other sensors. Most of these devices (44/46, 96%) employed an opportunistic approach to data collection, autonomously sensing and recording data without requiring users to manually input information or activate processes. The technical specifications of the WDs in each included study are detailed in Multimedia Appendices 5 and 6.

In the included studies, classification was the most commonly used problem-solving strategy (45/46, 98%; Table 3). A variety of AI methods were used, with convolutional neural networks (CNNs) being the most popular (17/46, 37%), followed by random forest (14/46, 30%) and support vector machines (12/46, 26%). As shown in Figure 3, the adoption trends of these methods evolved over the years. The majority of the reviewed studies utilized AI for diagnosis and screening (44/46, 96%), while only a few focused on using wearable AI to predict sleep problems before they occurred (6/46, 13%). Approximately 31 studies reported a mean data set size of 59,647.4 (SD 133,284), with a range of 12-561,480. Open-source data were used in a small number of studies (7/46, 15%), whereas the majority relied on closed-source data (39/46, 85%). All studies (46/46, 100%) collected data using wearable technology, while 4 (9%) also incorporated self-reported questionnaires, and 2 (4%) utilized nonwearable technology, such as cell phones. The most commonly used data types for model development included breathing-related metrics (eg, respiratory rate and respiratory effort; 25/46, 54%), heart rate–related metrics (eg, heart rate, heart rate variability, and interbeat interval; 22/46, 48%), and body movement (activity levels; 17/46, 37%). A total of 23 studies reported the number of features used, ranging from 2 to 10,500, with an average of 497.7 (SD 2181).

The most commonly chosen reference standard was clinical assessment (35/46, 76%). As shown in Multimedia Appendix 7, PSG was the most frequently used clinical assessment method. To validate AI model performance, most studies used a train-test split and K-fold cross-validation (21/46, 46%). The most commonly used metric for evaluating AI algorithms was accuracy (34/46, 74%), followed by sensitivity (29/46, 63%) and specificity (27/46, 59%). Multimedia Appendices 8 and 9 provide details on AI model characteristics in each cited study.

This scoping review explored the features of wearable AI technology used for sleep disorders. Consistent with previous reviews [73,74], we observed a positive trend in the adoption of wearable technology, reflecting a growing interest in sleep disorder research. The majority of studies were conducted in Asia (19/46, 41%), nearly twice as many as those in North America (12/46, 26%) and Europe (13/46, 28%). The gap between Asia and North America-Europe may be multifaceted. Contributing factors could include regional differences in sleep patterns [75-79] and the availability and affordability of WDs in Asia. Most studies focused on the middle-aged population (mean age 45 years), reflecting the higher prevalence of sleep disorders such as insomnia, sleep apnea, and restless leg syndrome in this group [80].

Key findings include the dominance of wearable AI in sleep apnea research (39/46, 85%). This can be attributed to the high prevalence of sleep apnea [81], its detrimental health effects [82], the limitations of existing diagnostic techniques [83], and advancements in wearable technology, which have made sleep apnea a primary focus for the development of innovative wearable monitoring systems [25]. Commercially available WDs (30/46, 65%) were predominantly used due to their accessibility, affordability, and ease of use [84,85], reflecting a shift away from prototypes seen in previous studies [86]. Wrist-worn devices and accelerometer sensors were the most commonly utilized technologies (34/46, 74%), often combined with photoplethysmography sensors to enhance sleep staging accuracy [87]. Another key finding of this review is that, despite the availability of well-known sleep wearables such as the Actiwatch and Belun Ring, relatively few studies used these devices. This may be due to their high cost or their specialized design and marketing for specific sleep disorders.

More than two-thirds of the studies used AI for sleep disorder screening and diagnosis, highlighting its value as a diagnostic tool due to its scalability, ability to identify high-risk individuals, and capacity to detect sleep disorders from wearable sensor data [88,89]. CNNs were the most commonly used AI models (17/46, 37%), likely due to the nature of wearable data, which are collected from raw sensors and require extensive preprocessing, including feature engineering and data cleaning. CNNs are well-suited for this task as they excel in handling complex data, extracting key features, modeling nonlinear relationships, and performing effectively on large data sets [90]. As shown in Multimedia Appendix 7, from 2013 to 2023, there has been a growing diversity in AI algorithms, with CNN, long short-term memory, and random forest remaining the most commonly used. The increasing adoption of ensemble and hybrid AI methods suggests a trend toward enhancing model performance. Data for AI models were predominantly sourced from closed data sets (39/46, 85%), with studies either recruiting their own participants or utilizing precollected hospital data. This preference may stem from privacy and ethical concerns, as these data sets often contain sensitive personal and physiological information, requiring additional safeguards and regulatory compliance for public sharing. The primary data types used for AI model development included respiratory data (25/46, 54%), heart rate (22/46, 48%), and body movement (17/46, 37%), as these are crucial for identifying the underlying causes of sleep disorders [43,91,92]. Respiratory rate was frequently utilized due to its critical role in detecting sleep apnea, the primary focus of most studies. Clinical assessments, particularly PSG (28/46, 61%), were the most commonly used reference standards for validation. PSG involves placing multiple sensors to monitor brain and heart activity, eye movements, muscle activity, blood oxygen levels, breathing patterns, body movements, snoring, and other noises, making it a widely preferred method for its accuracy and comprehensive assessment in sleep studies. Half of the studies validated their AI models using train-test split and K-fold cross-validation methods. K-fold cross-validation is especially effective at capturing data variability and is well-suited for smaller data sets, which are common in wearable studies [93]. However, the train-test split method was equally utilized. This preference may stem from its simplicity, ease of implementation, unbiased performance estimation, flexibility with data set size, and alignment with established best practices.

Our findings align with previous reviews [73,74], which reported an increasing use of wearable technology in sleep disorders research. However, unlike prior reviews that highlighted a focus on prototypes [86], we observed a significant shift toward commercially available devices, driven by technological advancements and affordability. Consistent with earlier studies [18,94], wrist-worn devices were the most commonly used, likely due to their portability and cost-effectiveness. While accelerometer-based wearables remained prevalent [95], this review highlights an emerging trend of integrating additional sensors, such as photoplethysmography, to enhance accuracy—an aspect less evident in earlier reviews. Furthermore, the increasing adoption of ensemble and hybrid AI methods represents a recent development in wearable AI applications for sleep disorders.

This review comprehensively assessed wearable AI technologies for sleep disorders, offering insights into their applications, regional trends, and preferences for sensors and algorithms. A key strength of this review is its focus on noninvasive WDs deployed in studies. By including research spanning a decade (2013-2023), we captured evolving trends in wearable technology and AI methodologies. Additionally, an extensive search across 7 diverse databases (eg, MEDLINE, Embase, IEEE Xplore) encompassing psychological, biomedical, technological, and interdisciplinary research ensured a comprehensive analysis.

First, this scoping review focused solely on WDs worn on the body, excluding nonwearable, implanted, and handheld devices (such as smartphones and carry-on sensors); near-body sensors (eg, Bluetooth transmitters); and devices requiring clinical intervention. As a result, the generalizability of our findings to such devices may be limited. However, by narrowing the scope, we ensured a focused review of wearable AI applications that are accessible and user-friendly. Second, we excluded studies that examined AI applications for detecting sleep quality or sleep staging without directly addressing sleep disorders, as well as those forecasting the outcomes of interventions for sleep disorders. Future reviews could broaden the scope to include these areas, providing a more holistic understanding of wearable AI applications in sleep research. Third, only studies published in English were included, which may have led to the omission of relevant research in other languages. Fourth, this review focused solely on the features of WDs and AI models and did not evaluate the efficacy or performance of wearable AI, as this was beyond its scope. Systematic reviews and meta-analyses, which assess quality and validate performance, are needed for such evaluations. Fifth, the rapidly evolving nature of wearable AI technology may mean that some recent advancements were not captured. Frequent updates to scoping reviews and systematic reviews can ensure timely insights into this dynamic field.

Noninvasive wearable AI devices hold significant potential for detecting and monitoring sleep disorders. Our review highlights a growing global research trend in this area. However, to comprehensively assess the performance of wearable AI, further systematic reviews are needed to statistically synthesize study results. Additionally, more research should explore wearable AI applications beyond sleep apnea. Future AI developments should extend beyond diagnosis and screening to include predicting sleep disorders, delivering personalized interventions, and providing tailored recommendations. Advanced AI models, such as generative AI and LLMs, should be explored in line with current technological trends. Manufacturers should integrate these models into WDs to enhance functionality and user experience. Additionally, studies should provide sufficient details on findings and model architectures to facilitate comprehensive systematic reviews and meta-analyses.