Given the relative novelty of this area of research, a scoping review was undertaken to provide an overview of the literature. This review adopted the 5-stage framework by Arksey and O’Malley [15] and reported according to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Scoping Reviews) guidelines [16], which is as follows: (1) identifying the research question; (2) identifying relevant studies; (3) selecting the studies; (4) charting the data; and (5) collating, summarizing, and reporting the results. As a scoping review aims to map the extent and nature of studies in the literature rather assessing their quality [15], the team did not perform a formal quality appraisal on the studies thus included.

The main research question formulated to guide this review was as follows: “How have machine learning technologies been used as a strategy for health promotion and behavioral change interventions?”

A search strategy was formulated to identify studies on ML techniques to promote behavioral changes and physical or mental health. A total of 9 databases were searched from inception to February 2021, aimed at encompassing not only multiple disciplines (Scopus and Web of Science), but also specific disciplines, including biomedical (the Cochrane Library, Embase, PubMed, Ovid, and ProQuest), nursing and allied health (CINAHL), and psychology (PsycInfo). To maximize the number of articles generated, no limits were applied to the dates and types of publications in order to maintain a comprehensive and updated search [17]. Relevant keywords and Medical Subject Headings (MeSH) terms were used, including “Learning, Machine,” “Behavior Control,” “Healthcare,” and “Mental Health” (Multimedia Appendix 1). Where appropriate, the keywords were truncated, and Boolean terms were added to maximize the retrieval of all relevant articles. End references of the studies were also hand searched for relevant articles [18].

Primary studies (including user preliminary testing or preliminary studies) were eligible for title and abstract screening if they had incorporated ML in promoting health or behavioral changes and were published in English (since the team lacked access to interpreters). Since this review aimed to examine real-life implications and experiences of incorporating ML techniques in health applications, articles were included only if they had studied at least a group of participants. Accordingly, those with no such reported results such as conference abstracts, proposals, and newspaper columns were excluded. Additionally, articles were excluded if they had examined other areas of health, such as the detection or diagnosis of conditions, population-health monitoring, prognosis and prediction of treatment outcomes, and research and clinical administration. Lastly, as ML techniques might overlap with statistical approaches [19], the 3 reviewers exercised discretion in determining those articles that had deployed ML techniques for inclusion.

All records retrieved from database searches were uploaded onto EndNote X9 (Clarivate Analytics), followed by the electronic removal of duplicates. During the screening of titles and abstracts, the remaining articles were independently reviewed by 2 reviewers (JOY and BC). Articles not meeting the eligibility criteria were removed, while those deemed suitable by at least one reviewer were downloaded for further examination. The opinion of a third reviewer (YSG) was sought, and disagreements between the reviewers were consensually resolved.

For each study, data extraction was conducted by one reviewer (JOY) and independently verified by a second (YSG). The following information was tabulated: author, year of publication, aim, main findings, area of health promotion, user data analyzed, type of ML used, challenges encountered, and future research. Such information served to inform the subsequent formation of themes in this review.

The literature search concluded in February 2021 and yielded 2673 search results. The removal of 1035 duplicates was followed by the screening of the titles and abstracts, during which another 1481 articles were removed. Of the remaining 157 articles included in full-text screening, 128 (82%) were removed with reasons such as duplicated articles, not using ML as main components, or not focusing on health behavioral change, leaving 29 (18.4%) for inclusion in this review (Figure 1).

The included articles were published between 2001 and 2020. The domains of health promotion presented in the articles were diverse, which are as follows: physical activity (9/29, 31%); smoking cessation (4/29, 14%); physical activity and dietary habits (4/29, 14%); stress detection or management (2/29, 7%); weight management (2/29, 7%); physical activity and sun-protection behavior (1/29, 3%); general self-health management (1/29, 3%); self-management for mental health (1/29, 3%); self-health management for hypertension (1/29, 3%); self-health management for sickle cell disease (1/29, 3%); help-seeking behavior for heart attack (1/29, 3%); diabetes prevention (1/29, 3%); and mindfulness meditation (1/29, 3%; Multimedia Appendix 2). Three themes emerged from the findings, which are as follows: (1) enablers, which it is the adoption of information technology for optimizing systemic operation; (2) challenges, which are the discussion of various hurdles and limitations presented in the articles; and (3) future directions, which explore prospective strategies in health promotion through ML.

This review has provided an insightful overview of the use of ML technologies in health promotion and behavioral changes, alongside a discussion of the challenges and potential opportunities for future works. Barring 4 studies [22,37,43,47], most others have been conducted over the past decade, indicating a recent growing interest in this topic. Additionally, the majority of the interventions in this review involved physical activity, whereas those specific to certain illness or those targeting smoking or mental health were relatively scarce. Hence, such interventions involving physical activity may be of interest for future works.

One noteworthy finding across the reviewed studies was the amount of time required to develop their ML-based applications. Some of the studies [20-25] specified the methods employed, including supervised ML, unsupervised ML, and reinforcement ML techniques. These differed in their use of supervised (labelled) data sets, which needed additional human intervention, and unsupervised (unlabeled) data sets. Unsupervised ML, coupled with the incorporation of virtual agents, embodied conversational agents, and human-like robots, required substantial time and resources to program natural language processing and to train the ML models [27,40,42]. Given the need for such length of time, the methodological implication was that most of the reviewed studies involved prototype research and pilot studies.

Furthermore, the complexity and comprehensiveness of ML-based applications grew with increasing amounts of data fed to their algorithms. The practical implication was that the experience differed between early-stage participants (who might experience a more rudimentary system) and their late-stage counterparts (who experienced a more refined one) [41]. These findings indicated that, to appraise the impacts of such ML-based applications more prudently and confidently, more studies examining them in a later, more mature, developmental stage would be warranted. Accordingly, developers might consider structured data input such as multiple-choice–based options [40] to aid the ML systems during the training of the application to refine and improve the accuracy of the feedback. These unique aspects of ML-based applications would deserve consideration in future studies incorporating ML techniques in their designs.

Another notable finding concerned hurdles in the design of the studies that might compromise their results, namely the lack of control groups [39,43,44,48], limited periods or samples for data collection [21-23,29,33,35,38,39,41,43-48], and potential confounders such as participation in other similar interventions [39]. To improve study designs and minimize the risk of confounders, the deployment of a waiting list control group and the disclosure of participation in other similar interventions should be instituted as part of the study methodology. Additionally, for short-term pilot and prototype studies with limited data collection, follow-up longitudinal studies might provide not only more data that would benefit the training of the ML system, but also insights into potential lasting effects of the intervention. Lastly, against such a background of these hurdles, health technologies would benefit from a set of guidelines and standardization, as advocated by Kulyk et al [31], for constant evaluation at all developmental stages to maximize the users’ potential to meet the target health behaviors.

Two other aspects merited discussion, which are as follows: (1) the burden on users to provide accurate data for tracking their lifestyles through self-reporting and the use of sensors; and (2) the challenges surrounding personalization. Although systems detecting sensor failures [20] and erroneous self-reported measures [34] might enhance the accuracy of behavioral measurement outcomes, further work in this area of research would be needed, especially with the increase in sensor usage in recent years [50]. Moreover, methods to reduce sensor failures through hardware enhancements could improve the quality of data. Additionally, both under- and overpersonalization would lead to overly generic feedback for users and breaches of privacy; this aspect thus warranted developers’ prudent and sensitive handling to strike a balance between providing customization and upholding privacy. Taken together, the challenges identified by the reviewed studies have underscored the need for a design that is nonobtrusive, requires minimal users’ input, encourages continuous engagement, and enables personalization while not encroaching upon privacy. Alongside such efforts, greater investment in combining multiple ML and AI techniques for synergy should likewise be pursued.

This review offers a timely overview on the emerging field of ML in health technologies, alongside potential referential value for future works. Nonetheless, several limitations are noteworthy. Firstly, the findings may not be generalizable since this review has included studies in only English due to the lack of access to interpreters. Additionally, the predefined scope of this review meant that studies examining other areas of health, such as detecting and diagnosing medical conditions, were excluded. Secondly, despite the use of numerous combinations to attempt to encapsulate the concept of ML in our search strategy for this review, the wide variety of ML techniques, alongside their overlap with statistical approaches [19], might have resulted in an omission of eligible studies. Finally, despite the demonstrable viability of ML techniques in health promotion and behavioral changes and their potential for integration into health care systems, most of the reviewed works have been pilot studies or prototype research, for which more investigations would thus be warranted to determine the clinical utility of the interventions.

This review has examined the use of ML in health promotion and behavioral changes by mapping its associated challenges and highlighting potential areas for future works. The findings have collectively demonstrated that the challenges pertained to not only the time- and resource-consuming nature of ML-based applications, but also problems such as the burden on users for data input and the degree of personalization. Future works may consider designs that correspondingly mitigate these challenges in areas such as mental health promotion where the use of ML remains limited.