Overview of Artificial Intelligence–Driven Wearable Devices for Diabetes: Scoping Review

Background Prevalence of diabetes has steadily increased over the last few decades with 1.5 million deaths reported in 2012 alone. Traditionally, analyzing patients with diabetes has remained a largely invasive approach. Wearable devices (WDs) make use of sensors historically reserved for hospital settings. WDs coupled with artificial intelligence (AI) algorithms show promise to help understand and conclude meaningful information from the gathered data and provide advanced and clinically meaningful analytics. Objective This review aimed to provide an overview of AI-driven WD features for diabetes and their use in monitoring diabetes-related parameters. Methods We searched 7 of the most popular bibliographic databases using 3 groups of search terms related to diabetes, WDs, and AI. A 2-stage process was followed for study selection: reading abstracts and titles followed by full-text screening. Two reviewers independently performed study selection and data extraction, and disagreements were resolved by consensus. A narrative approach was used to synthesize the data. Results From an initial 3872 studies, we report the features from 37 studies post filtering according to our predefined inclusion criteria. Most of the studies targeted type 1 diabetes, type 2 diabetes, or both (21/37, 57%). Many studies (15/37, 41%) reported blood glucose as their main measurement. More than half of the studies (21/37, 57%) had the aim of estimation and prediction of glucose or glucose level monitoring. Over half of the reviewed studies looked at wrist-worn devices. Only 41% of the study devices were commercially available. We observed the use of multiple sensors with photoplethysmography sensors being most prevalent in 32% (12/37) of studies. Studies reported and compared >1 machine learning (ML) model with high levels of accuracy. Support vector machine was the most reported (13/37, 35%), followed by random forest (12/37, 32%). Conclusions This review is the most extensive work, to date, summarizing WDs that use ML for people with diabetes, and provides research direction to those wanting to further contribute to this emerging field. Given the advancements in WD technologies replacing the need for invasive hospital setting devices, we see great advancement potential in this domain. Further work is needed to validate the ML approaches on clinical data from WDs and provide meaningful analytics that could serve as data gathering, monitoring, prediction, classification, and recommendation devices in the context of diabetes.


Introduction
Background Diabetes, also known as diabetes mellitus, is a metabolic disease characterized by elevated blood glucose levels, which can ultimately result in many complications such as heart attack, stroke, kidney failure, leg amputation, vision loss, and nerve damage [1]. As the world embarks on a centennial anniversary since the development of insulin to manage glucose levels of people with diabetes, we have seen remarkable advances during these 100 years, with improved life expectancy and quality of life [2]. Noncommunicable diseases such as metabolic syndrome and diabetes continue to be among the leading causes of disability and mortality [3]. The number of cases and their prevalence have steadily increased over the last few decades. According to the World Health Organization, 1.5 million people died in 2012 alone because of diabetes, with an additional 2.1 million deaths caused by a higher than optimal blood glucose level, resulting in increased risks of cardiovascular and other diseases. A total of 463 million people, globally, were affected by type 2 diabetes (T2D) mellitus in 2019. Furthermore, it is predicted that 700 million individuals would develop diabetes by 2045 [4]. Although the World Health Organization acknowledges that there is no one fixed solution and that a coordinated multicomponent intervention is needed, it outlines technology as one of the key stakeholders in reducing the impact of diabetes in addition to input from governments, health care providers, people with diabetes, civil society, food producers and manufacturers, and suppliers of medicine [1].
Despite the advancements in blood glucose monitoring techniques, the mainstream detection technology remains largely invasive. The commonly used home electronic glucometers involve people with diabetes invasively self-pricking to draw blood from fingertips, opening them up to infections as well as stress and pain caused by the procedure that is often expected multiple times a day.
The availability and advancements of smart devices, such as smartphones, have made the monitoring of diabetes-related features more accessible. Many studies have examined this much welcomed technology [5,6]. These normally require the use of an external attachable sensor, and monitoring is then delivered via an app or a separate continuous glucose monitoring (CGM) device, which can still be semi-invasive and require a connection range via Bluetooth or Wi-Fi signals. The use of completely noninvasive technology in the form of wearable devices (WDs) for regulating and monitoring glucose levels for people with diabetes is a fairly new concept and is in its infancy.
Commercially available devices, such as smart watches and smart bands, can take measurements using sensors that researchers have reported on their usefulness in diabetes monitoring [7,8]. Such technologies can be affordable and easily accessible, and when used properly, can improve the quality of life of patients in a noninvasive manner. With their widespread commercial use and acceptance owing to their fashionable nature, globally researchers have a unique opportunity to provide medical care away from hospital settings and bulky invasive hardware in an affordable manner without requiring expert assistance. WDs have an increasing capacity, although not at the level of smartphones, to gather, store, transmit, and process data; the features can then be used for management, treatment, assessment, and sometimes even prediction. Furthermore, many WDs are normally connected via Wi-Fi or Bluetooth to external devices, such as a smartphone, where computationally expensive processing is performed for the simple purpose of storage or as a gateway to cloud spaces. Cloud storage can facilitate monitoring by clinicians without the need of hospitalization. Several useful sensors already exist incorporated into WDs similar to those of smartphones, including electrocardiogram (ECG), photoplethysmography, galvanic skin response, near infrared, and accelerometer sensors. WDs have additional advantages when it comes to sensing physiological signs, such as heart rate, ECG, and skin temperature. This is largely owing to their close contact with the wearer, which is of particular interest when monitoring diabetes-related metrics.
Artificial intelligence (AI) is a broader term that encompasses machine learning (ML). Technically, ML is a subset of AI, often loosely used interchangeable buzzwords. As a high-level definition, AI is anything related to making machines smarter (eg, computational search algorithms). ML, on the other hand, is an AI system that can self-learn via an algorithm, and as a result, such a system becomes smarter without human intervention over time (eg, classifying an outcome) [9]. Deep learning, on the other hand, is another branch of AI that attempts to mimic the human brain in terms of how it processes large amounts of data and has already shown success rates in areas such as diabetic retinopathy screening [10]. ML principles have been applied in clinical settings to build algorithms to support predictive models for the risk of development of diabetes [11]. AI has also been shown to provide useful management tools to deal with large amounts of data [12]. Owing to the large amount of data measurable through continuous monitoring via wearables, AI can be used to further analyze the acquired data. This can help to understand and draw meaningful information from the gathered data and provide advanced and clinically meaningful analytics. Many researchers have adapted existing WDs not originally intended for diabetes management and adapted the sensory information for use in diabetes-related metrics, and some have created prototypes especially designed for diabetes [13,14]. WDs are used for a variety of reasons, including monitoring, prevention, glucose estimation, diagnostics, classification, and prevention, but the number of studies that are reported are low in comparison with those that make use of smartphones for example. With the increased potential outreach of WDs globally, especially when combined with the ever-expanding field of AI-incorporating ML algorithms, the correct management of large amounts of data and processing with ML algorithms holds great potential for quality-of-life improvement in people with diabetes [15].

Research Problem and Aim
Many studies have been conducted on AI-based WDs for diabetes. Exploring the features of AI-based WDs reported in these studies is important for developers, patients, health care providers, and researchers to identify the recent advances and challenges in this area. Although several reviews were conducted in this area, (1) they were focused on smartphones and AI for diabetes [16][17][18], (2) they were focused on WDs in general rather than AI-based WDs [17,19], and (3) they did not summarize the features of AI-based WDs in a thorough manner [16][17][18][19]. Therefore, we aimed to explore the features of AI-based WDs for diabetes as reported in previous studies. We believe that this review will allow developers and researchers to advance further in this field by highlighting the gaps and opportunities.

Overview
This scoping review was carried out to satisfy this study's goals of exploring features of AI-driven wearable technologies for diabetes. In order to construct a complete scoping review, the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [20] was used as a guiding approach. The PRISMA-ScR checklist is shown in Multimedia Appendix 1.

Search Sources
The article search for this review began by identifying all relevant studies using 7 electronic databases: MEDLINE, PsycINFO, EMBASE, IEEE Xplore, ACM Digital Library, Web of Science, and Google Scholar. We scanned the first 100 hits retrieved by searching Google Scholar. The reason being Google Scholar typically returns several items that are sorted by relevance to the search topic. Bibliographic collection was conducted from October 25 to October 30, 2021. The reference lists of the included articles were then searched for additional sources. We also checked relevant articles that cited the included studies using Google Scholar's "cited by" tool (forward reference list checking).

Search Terms
A number of different sets of keywords were designed to search databases depending on each database's search term limit; as IEEE and Google Scholar have term limits, search queries were truncated based on the required limit. We considered the research topics included in the database to complete our search queries. We combined Diabetic OR Diabetes keywords describing the relevant population (people with diabetes), with each kind of relevant intervention to wearables (wearable* OR smart watch* OR smart* OR smartwatch* OR fitness band* OR flexible band* OR wristband* OR smart insole* OR bracelet*) and AI ( Studies were chosen based on the criteria in Textbox 1. Peer-reviewed articles and published protocols were included only if they were related to wearables that could be used by an individual outside of a clinical setting. They also had to use AI for the purpose of diabetes and be classified as noninvasive. For full inclusion and exclusion criteria refer to Textbox 1.  6. Wearables using methods for diabetes analysis are to be noninvasive.

Exclusion criteria
1. Any study that does not contain AI as an intervention.
2. People with other diseases, health care providers, and caregivers as population.
3. Not a wearable device (example artificial implant or body infused). 4. Studies opting statistical measures only, for analysis of collected data.
5. Sensors or tracking devices infused inside a person's body. 6. Wearable devices that need professional sittings or hospital sittings.

Study Selection
This review's studies were selected in 2 steps. In the first stage, 2 reviewers (AA and SA) independently reviewed the titles and abstracts of all retrieved papers. In the second phase, the same reviewers individually read the whole texts of the papers included in the first step. Rayyan (Qatar Computing Research Institute, Hamad Bin Khalifa University) [21], a web-based tool developed for data management for systematic and scoping reviews, was used to upload all the articles acquired from databases in a Research Information Systems format; then, filtering and citations were managed. During the first and second steps of the selection process, any disagreements between the 2 reviewers were resolved through conversation and decisions were made based on consensus.

Data Extraction
AA and SA constructed the data extraction form, as shown in Multimedia Appendix 3. The data extraction technique was carried out independently by 2 reviewers (AA and SA), and any discrepancies were resolved by discussion and consensus. Microsoft Excel was used to record the data extracted.

Data Synthesis
SA synthesized the extracted data using the narrative approach, aggregating the data using tables and text and nonstatistical techniques. For being more precise, we presented the search results followed by general features of the studies, finally describing characteristics of the WDs and AI technologies. We described the general features of WDs (eg, device placement, type, and operating system [OS]) and their technical features (ie, features of sensors, such as sensors used, sensing approach, and primary measurements). The AI features were addressed based on the models used, the evaluation metrics, and their applications.

Search Results
Having searched 7 bibliographic databases, this study returned 3872 citations. As shown in Figure 1, a total of 294 duplicates were subsequently removed, leaving 3578 unique titles, and abstracts; publications that did not make use of AI technologies via WDs for diabetes management were considered irrelevant. Of these, we further excluded 3424 citations after screening their titles and abstracts. Of the remaining 154 references, 117 publications were excluded during the full-text screening. We were left with 37 studies, and this number remained unchanged even after performing backward and forward reference list checking. The synthesis included a total of 37 articles (Multimedia Appendix 4 [7,8,13,14,).   (21/37, 57%). A large proportion of the studies were authored by institutes in the United States (7/37, 19%), China (5/37, 14%), and India (5/37, 14%). A total of 26 of the 37 studies (70%) were journal articles, and the remainder were conference proceedings (11/37, 30%). Most of the studies targeted type 1 diabetes (T1D), T2D, or both (21/37, 57%), whereas 32% (12/37) did not specify the type of diabetes and mentioned diabetes in general. The remainder targeted prediabetes or a combination of T1D, T2D, and prediabetes (4/37, 11%). Features of each included study are shown in Multimedia Appendix 5.

Technical Features of Wearables
Most of the studies (28/37, 76%) reported an opportunistic approach (ie, no input required from the participant) when obtaining data using the WDs, whereas the remaining (9/37, 24%) used a participatory approach (ie, input required from the participants). For sensing technologies, various sensors were used, either built-in to the WD or as wearable sensors, often reported as >1 sensor per device. We observed a large number of devices in the studies reviewed reporting photoplethysmography sensor use (12/37, 32%), while optical heart rate was only seen in 5% (2/37) of studies among some of the other less-reported sensors. Features of WDs for each included study are shown in Multimedia Appendix 5.

Wearable Technology Status Versus WD Type
Multimedia Appendix 6 further visualizes the data highlighting the WD type and whether they are commercial or prototypes. Wearable sensors were the most prominent as a prototype while smartwatches and smart wristbands were the most common as commercial. Figure 2 shows the type of diabetes and number of studies related to each WD type. While most studies did not specify the type, T1D (as a smart wristband), T2D (as a wearable sensor), or both (as a wearable sensor or smartwatch) seem to be the most targeted types.

AI and ML Technologies
For the purpose of this study, we categorized the ML algorithms into 4 categories (classification models, regression models, neural network-based models, and optimization algorithms) and those that were not clearly specified by the study authors were categorized as black boxes (ie, studies that mention they make use of ML or AI but do not specify any further details of algorithms used). Many ML technologies were reported that come under these headings (refer to Table 5 for a full list), and some studies reported and compared >1 model. Support vector machine (SVM) was the most reported (13/37, 35%), followed by random forest (12/37, 32%), k-nearest neighbor (7/37, 19%), Naive Bayes (5/37, 14%), and decision trees (4/37, 11%) among the most used models from classification models. From the regression models, only linear regression (2/37, 5%) was reported in a couple of studies, whereas all others were reported by single studies only. Artificial neural networks were reported in 14% (5/37) of the studies in neural network-based models, followed by long short-term memory (4/37, 11%), convolutional neural networks (3/37, 8%), and deep neural networks (3/37, 8%); these networks were used for both classification and regression purposes. Table 5 also highlights that the majority of the studies applied the AI and ML technologies for either the purpose of blood glucose level forecasting (12/37, 32%) or classifying the participants as normal, diabetic, or prediabetic (12/37, 32%). Table 6 highlights some of the statistical measures used to evaluate the ML algorithms within the reported studies. Some studies used multiple statistical techniques for this purpose, among them were reports of accuracy (20/37, 54%) and sensitivity (9/37, 24%). While some studies did not mention which was the best ML model identified (6/37, 16%), random forest was reported as the best identified model (7/37, 19%), followed by SVM (6/37, 16%). Table 5. Artificial intelligence (AI)-and machine learning (ML)-related features (n=37).

Principal Findings
This was the first study of its kind to the best of our knowledge, considering the amount of features we were able to extract from each publication. The features extracted should give researchers insight not only into the technologies that are readily available commercially but also into what is possible in the future with studies we identified that developed prototypes. Our findings shed light on this emerging field, which is still in its infancy. This is further highlighted by the fact that 59% (22/37) of the studies that met our inclusion criteria were prototypes; we were only able to identify 41% as commercially available (as demonstrated in Multimedia Appendix 6) devices, of which only (7/15, 46%) studies performed some sort of ML classification on the extracted data directly from WDs, whereas (6/15, 40%) studies made use of neural network-based models with classification to make out of already collected data. Most of these measured blood glucose on wrist-worn devices and used a classification algorithm (Figure 3). Classification models were widely used (Figure 3) in the reviewed studies, largely owing to studies attempting to classify types of diabetes (T1D, T2D, etc). SVM and random forest were the most prevalent classifiers and exhibited the highest performance. SVM [55] is extensively used because of its superiority in generalization and nonlinear function fitting, and it also has a number of advantages when dealing with small-sample studies [56]. Furthermore, SVM is a binary classifier, and we observed that it is mostly used on blood glucose level data to determine levels for diabetes categorization. Aside from the accuracy for demonstrating efficacy, the Clarke Error Grid was the most commonly used performance metric, possibly because of its popularity as a performance metric for assessing blood glucose estimation. The grid was split into 5 zones, each with varied prediction accuracy between the estimated and reference blood glucose readings. The data fell within zone A, which pertains to precise glucose calculations, where each consecutive zone is thought to have progressively substantial erroneous estimations [57,58]. Most of the sensory data being collected especially when looking at commercially available devices, did not require any further or minimal input from the user, meaning the person with diabetes can get on with day-to-day tasks without having to worry about taking regular invasive finger pricks for monitoring glucose levels; for example, while still feeling that they are wearing a stylish item such as a smartwatch. We specifically examined studies after 2015, as previous studies related to the use of WDs found that most wearables were used in this range [59]. One of the reasons may be that Fitbit released its first device in 2009 and the Apple Watch followed in 2015; both these devices set the tone for WDs, and it is not surprising that 59% (22/37) from our review were wrist-worn. A total of 78% (29/37) of the devices were connected to either a gateway or host device, usually a smartphone (16/37, 43%) via either Bluetooth (19/37, 51%) or Wi-Fi or internet (6/37, 16%); this is likely owing to the fact that web-based data are now more affordable and the availability of low energy connectivity technology such as Bluetooth. This ability to connect has resulted in more analytics and data storage being possible on host devices than on smartphones or directly on the cloud (18/37, 49%) of studies in this review, compared with limited computing power on the WD itself. One of the limitations of this is that devices need continuous connections, which can be an issue, as reported data can be lost if the connection is not maintained for long periods [60]. We also observed that many devices used gateways or host devices, which we believe to be largely because of the limited computing power of WDs.

Strengths
This review was conducted according to the PRISMA-ScR; therefore, it can be considered a high standard. Two reviewers independently conducted the study selection and data extraction. We believe this to be the first of its kind study focusing on WDs targeting diabetes using AI approaches and were unable to identify previous scoping reviews in the literature that has as an exhaustive list of features extracted in this field. A combination of expert research computer scientists and research medical practitioners allowed us to explore the current technologies in depth and highlight gaps in the research community. The most popular databases in the health care and information technology fields were searched; furthermore, Google Scholar with forward and backward reference list checking allowed an exhaustive search of the literature, reducing the risk of publication bias.

Limitations
Only studies published between 2015 and 2021 in the English language were included. Furthermore, we did not use Medical Subject Headings terms in our search; therefore, we may have overlooked some relevant studies. We excluded devices that could be classified as WDs, such as electroencephalogram and ECG machines, which limited their use in hospital settings. As our focus was AI, we excluded any study of WDs and diabetes that had a statistical measurement not considered an AI approach. Although we included a large number of features and some effectiveness measures, we fall short of critically assessing the quality of each of the included studies-this goes beyond the scope of our review-and we hope to cover this in a full systematic review in the near future on the same topic.

Practical and Research Implications
WDs hold great potential for the self-monitoring of diabetes-related parameters, and their ability to be paired with a range of smart devices, including smartphones and general connectivity to clouds, allows the continuous collection of data from many biosensors that measure vitals and biosignals without user interference. The fact that they can be worn in a stylish and fashionable manner has potential for wider acceptance than other technologies, such as CGMs. Although many studies have used WDs for diabetes, we found that ML is still lacking in a sizable number of these studies. With the limited number of studies that reported the use of ML, we see great promise, largely owing to the accuracy levels of the ML algorithms reported in Table 6. Engineering and data science research experts need to come together and identify the most common sensors and technologies and study their effectiveness when combined with ML approaches. In addition, commercially available WDs are readily available and therefore sit in waiting for researchers to conduct studies and apply ML and report further in scientific journals to prove validity and instill consumer confidence. Most of the papers identified in this study used AI or ML algorithms for testing the validity of the system functioning rather than identifying the approaches that could be used for the development of such intelligent devices. More work needs to concentrate on applying known ML algorithms for the purpose of making more accurate diabetes-related measurement calculations. Currently, the number of commercial devices associated with studies are still very low, a quick search on retail sites such as Amazon reveals many commercial devices claiming diabetes-related measurements, which have still yet to be validated with related studies, and this is one area where researchers could get involved. Researchers need to make more use of purpose-shifted devices as they are lying in wait as opposed to creating prototypes and testing the effectiveness of the many commercially available devices. We encourage researchers to perform systematic reviews to assess the efficacy of AI-based and non-AI-based WDs compared with traditional medical devices. Some technologies that are classified as WDs such as CGMs are still classified as semi-invasive as they allow the embedding of a sensor partially into participants' skin, we feel for wider acceptance especially for home use products the technology really needs to move away from such sensors and more studies now need to focus on how measurements can be obtained from noninvasive sensors such as those available on commercial smart watches. Further work is also required on ML algorithms used for diabetes data that can be used on the WDs as opposed to on host devices, as this would reduce some issues reported such as loss of data owing to WD out of range with the host device, which will become easier with time as the technology advances and WD memories are no longer a limitation. We suspect there would be less reliance on host devices for some of the ML computations.
Another area for exploration is the use of the internet of things (IoT); in our search, we found a handful of studies making use of IoT. Most IoT papers describe the IoT architecture for diabetes management without specifying the sensors or WDs actually used or implemented, and do not go into much (if any) detail about any ML deployed. There are many opportunities in this domain; none of the studies were found to make good use of developed commercial technologies such as Alexa, Google Home, and Apple watches, which are readily available. The possibilities here are endless, using a combination of data gathered from sensors at the WDs with other patients and personal data in real-time with IoT. This brings along with its own caveats and the need to incorporate questions of privacy and data sovereignty arising from the mass data storage in cloud-based systems and the many interconnected devices and hospital datacenters; there are issues that need to be considered with the use of data and individual consent. There are also problems regarding the scope of an individual's consent to use their data, as well as potential accountability if the data are mishandled. There are dangers associated with AI algorithms and their misdiagnoses, dangerous advice, or recommendations that do not correspond to the required standard of care. Data security breaches or the reidentification of previously deidentified data may have unintended repercussions. Furthermore, other ethical issues need to be considered, such as accessibility, although commercial WDs that are easily and cheaply available may not be affordable for the masses in low-income countries. A multidisciplinary effort is required, including but not limited to engineers, medical practitioners, and legal experts.

Conclusions
We investigated and reported the current state of WDs and their features for the purpose of diabetes that use ML approaches. Considering the availability of consumer-grade biosensors, we see great advancement potential in this domain, replacing hospital setting, invasive devices, especially when it comes to monitoring glucose levels. Further clinically significant studies are needed to instill confidence and validate WD use as well as the application of ML algorithms on WD data. Researchers and those wanting to develop AI-based WDs can use our review to understand where the gaps are in this emerging field. We encourage readers to use more data and delve deeper into the studies we have identified in order to establish, validate, and repeat studies that showed high accuracy. There is still much work needed, and we feel our review has provided the most extensive work so far summarizing WDs that use ML for people with diabetes to date. Finally, researchers will also benefit from our study as they can embark on longer and better populated systematic studies scrutinizing the benefits of WDs as data gathering, monitoring, prediction, classification, and recommendation devices in the context of diabetes. We envisage several follow-up studies, starting with a full systematic review from our own group.