Artificial intelligence (AI) is a concept that dates back to the mid-1900s [1] and was first defined as “the science and engineering of making intelligent machines” [2]. Today, AI is described as a computer system’s capability to perform complex tasks that mimic human cognitive functions to perform tasks such as reasoning, decision-making, or problem-solving, autonomously and adaptively [3]. However, its capabilities and underlying functions have changed significantly over the decades [1,4]. More recently, AI has emerged as a transformative force across various industries. Its application has shown promise in health promotion and health care [5-7], opening new possibilities concerning patient care and enhanced medical practices.

There is growing consensus in the literature that adaptivity and autonomy are the key characteristics of AI applications and technologies [5]. AI is considered an umbrella concept of emerging technologies, enclosing fundamental distinct types such as machine learning (ML), deep learning (DL), or natural language processing (NLP) [4,8]. Technically, AI is an ML-based approach that simulates human minds’ cognitive and affective functions [3,8] and is designed to observe and react to a specific environment. In contrast to deterministic programming, such models feature many free parameters that can adapt autonomously to calibrate the model. For example, AI can be applied in repetitive tasks requiring human intelligence, such as scanning and interpreting magnetic resonance imaging, autonomous driving, or analyzing big data sets [9-11]. ML and DL algorithms and artificial neural networks enable a machine or system to learn from large data sets, make autonomous decisions, and improve their performance over time [4]. More narrowly, NLP allows machines to generate and understand text and spoken language in the same way humans do. It combines rule-based natural language modeling with ML and DL models to process human language in text or speech data, understand its meaning, including feelings, and even generate human language, as it is sometimes used in chatbots or language translation [12].

Implementing AI algorithms and technologies for health care institutions bears enormous potential, ranging from efficient health service management, predictive medicine, patient data, and diagnostics with real-time analyses to clinical decision-making. Most studies report a broader AI architecture with a combination of algorithms rooted in ML, DL, and NLP [4,11]. For example, 1 AI approach evaluated the support of clinical decision-making by analyzing continuous laboratory data, past clinical notes, and current information of physicians synthesizing significant associations [13]. AI implementation in the form of predictive modeling showed positive results by detecting irregular heartbeats through smartwatches [14], automatically identifying reports of infectious disease in the media [15], or ascertaining cardiovascular risk factors from retinal images [16]. Through systematic profiling of 4518 existing drugs against 578 cancer cell lines with an AI-based approach, a study revealed that nononcology drugs have an unexpectedly high rate of anticancer activity [17]. Another study developed and evaluated a Medical Instructed Real-Time Assistant that listens to the user’s chief complaint and predicts a specific disease [18]. Chatbots have been used to detect COVID-19 symptoms through detailed questioning [6] or to predict the risk of type II diabetes mellitus [19].

As adults spend a significant amount of time working, it is widely accepted that work and work environments have a major impact on individuals’ health. Workplace health promotion and prevention (WHPP) are important fields that “[…] improve the health and well-being of people at work […]” [20] through a combination of behavioral and organizational measures. Workplace health promotion follows a competence-oriented, salutogenetic approach to promoting the resources of an individual [20]. Prevention in the workplace focuses on minimizing the burden of workplace-related diseases and associated risk factors [21,22]. WHPP interventions range from behavioral measures with active participation (eg, courses or seminars) to organizational measures such as consultations, analyses, inspections, and establishing organizational structures such as a health committee [23,24].

With the Luxembourg declaration, WHPP has evolved into an independent discipline that differentiates from return-to-work (RTW) and occupational safety and health (OSH) measures [20,25]. In OSH-related disciplines, previous reviews have focused on risk assessment or detection related to physical ergonomics [26], occupational physical fatigue [27], or core body temperature [28]. Other reviews explored the evidence of AI in F-related areas, such as vocational rehabilitation [29] and functional capacity evaluation [30]. In health promotion in general, 1 review evaluates the use of chatbots to increase health-related behavior but does not focus on the workplace setting [31]. To the authors’ knowledge, no review has evaluated the use of AI in WHPP.

Therefore, this systematic scoping review aims to comprehensively assess an overview of the current use of AI in WHPP. The results will then be used to point to future research directions. The following research questions (RQ) were derived from these aims:

A systematic scoping review approach [32] was selected following the extended PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews; Multimedia Appendix 1) [33]. We applied the 5-step framework to identify current or emerging research directions and provide an overview of research activities [34]. Additionally, the patient or population, intervention, comparison, and outcomes (PICO) framework [35] was used to specify the study’s objective, from the search string and data charting to more systematic discussion [36]. The review was registered prospectively in the Open Science Framework (OSF) on July 5, 2023. All files (protocol, search string, and search results) have been uploaded to the OSF profile and are publicly accessible [37].

Included studies needed to be (1) peer-reviewed, (2) written in English, and (3) focused on any AI-based algorithm or technology that (4) were conducted in the context of WHPP, or (5) an associated field (workplace prevention, occupational health, and workplace health) that applies to WHPP. The types of research considered were review types (systematic, scoping, or rapid), cross-sectional studies, and longitudinal studies.

Our conceptualization of AI included the concepts of “machine learning,” “deep learning,” and “natural language processing.” Our conceptualization of “workplace health promotion and prevention” followed a broader understanding comprising the setting (eg, “work,” “workplace,” or “in or at the workplace”), the target population (eg, “working adults” or “employees”) and the outcome dimension (eg, “health” or “health behavior”). The search period was limited to studies published since January 2000 and before July 31, 2023. During the review, the search was extended to December 20, 2023.

The systematic literature research was conducted in July 2023 in 3 databases: PubMed, IEEE Xplore, and Association for Computing Machinery. The search string included Boolean operators (“AND,” “OR,” and “NOT”) and search terms related to “artificial intelligence,” “workplace health promotion,” “health promotion,” and “workplace setting” (see supplementary files available at OSF profile [37]). Papers were managed with the software tool Rayyan, followed by a 2-stage screening process. First, 1 reviewer (ML) removed all duplicates. Second, 2 reviewers (ML and AL) screened all titles or abstracts and read full texts for eligibility criteria in a blinded procedure. Disagreement was resolved by either consensus of the 2 reviewers or by consultation of a third reviewer (IK).

In the first step, the study characteristics were extracted: first author (name and year), study design (eg, cross-sectional or randomized controlled trial), the primary type of AI algorithm and technology as referred to in the study (eg, AI, ML, DL, or NLP), and the frontend in which the AI-technology was implemented (eg, mobile app or web app). Second, the PICO framework [35] was applied to extract information about the target group (number of included participants/workplace context), the intervention approach, the comparison, and the reported outcomes of the study.

We used the extracted information from the study characteristics to answer RQ1 on current AI-based technologies applied in WHPP. For answering RQ2 and RQ3, we used the data extracted by the PICO framework. The information was then categorized within the results’ tables and summarized narratively.

The predefined search led to a total of 3317 results. The screening results revealed 478 duplicates, 712 records not meeting inclusion criteria (eg, publication type, language, or setting), 42 unique records, and 104 with missing information, leaving 1981 records for the title and abstract screening. The title and abstract screening excluded another 1761 records for not meeting inclusion criteria, leading to 220 records for full-text screening, of which one was inaccessible. After screening 219 full-text records, another 209 records were excluded. Finally, 10 studies remained in this systematic scoping review (the PRISMA-ScR flowchart is shown in Figure 1).

The results of the study characteristics are presented in Table 1. Regarding the study designs, 6 studies were cross-sectional studies [38-43], 3 were randomized controlled trials [44-46], and 1 was a quasi-controlled trial [47]. None of the studies explained data protection standards (security protocols, storage location or duration, or access of third parties) within the AI algorithms and technologies used. In most studies, white-collar workers were the intended target group [38,41,42,46], whereas, in 3 studies, white-collar and physical labor workers participated [40,45,47]. Further, 1 study evaluated AI-based technologies with physical labor workers [39], and another did not disclose any information about the type of work setting [44]. Information on sample characteristics was missing in 3 studies [40,41,44], little information was provided in 2 studies [38,44], and 4 studies offered sufficient information [39,42].

A comparison was used in different ways by 6 studies [40,42,44-47]. Further, 4 studies recruited a classic control group [39,44,46,47], 2 of which exposed the control group after a waiting period [44,46]. Another study compared their assessed data to external data thresholds [40], and 1 study compared assessed objective data with subjective data [42]. Regarding the outcome, all studies stated sufficient and significant results. Further, 1 study reported no changes in 1 of the 3 assessed outcomes [47].

AI algorithms and technologies were mainly used for preventive purposes in risk assessment (Table 1). Furthermore, 2 studies evaluated prediction models [39,42]. Additionally, 3 studies [44,46,47] targeted health behavior change using 3 different approaches ranging from a web app [44] and smartphone app [46] to social robot agents [47]. Further, 1 study [41] was categorized as an organizational health promotion approach. A major target indication was mental health, which was addressed in 4 studies [38,40,42,43]. In contrast, 1 study dealt with musculoskeletal disorders [45] and 1 on overall physical health and work-related factors [39].

The PICO category “intervention” did not apply to studies focusing on prevention since they did not evaluate an intervention [38-43]. Interventions were evaluated by 4 studies [44-47] with a duration of 12 weeks [44-46] and 8 weeks [47]. Within these 4 studies, 2 used chatbots as a primary AI application [45,46], 1 used a web application [44], and 1 used a social robot agent [47]. These 4 studies recruited a control group, of which 2 studies exposed the control group after a waiting period [44,46]. Regarding the outcome, all studies stated sufficient and significant results. The study of Lopes et al [47] reported no changes in 1 of the 3 assessed outcomes (Table 2).

As pointed out, current research is on AI in WHPP regarding quantity, fields of WHPP and its subdomains, and AI algorithms. Future research should center around major causes of sick leave, such as musculoskeletal disorders, mental health, respiratory conditions, and influenza [67]. Behavioral WHPP interventions should extend to all areas of health-related behavior, including nutrition, sleep, substance abuse, and stress management [24]. Further, setting-specific aspects of WHPP, such as intervention content, implementation strategies, user experience, design, algorithms, and the company’s size, need to be considered more specifically. So far, the studies have provided only moderate information on the job activities or the target groups. At the same time, workplaces and workers are diverse. The health of employees is influenced by numerous organizational and individual factors that must be further considered in the learning cycle of AI [77-79]. Regarding potential errors, existing AI algorithms must be validated with different target groups [59,80], emphasizing the need for longitudinal data and its impact on learning algorithms [81,82]. Beyond this, the technological diversity of the presented studies opens new possibilities for target group-specific or individualized interventions. Providing health information to chronically ill employees, migrants with different language skills, or individualizing health topics of varying age groups can be provided more effectively through AI to move beyond a “one size fits” all paradigm [83,84].

Outside of the objective’s scope, we identified 2 aspects that can improve future research. First, the included studies reported overall positive results regarding feasibility, significance, or accuracy, underlining the vast potential that AI technology harbors. However, the results must be interpreted cautiously as certain information in the primary studies was not provided, assessed, or available at the stages of the investigated technology. For example, few studies mentioned a potential bias through the novelty [40,47] or the Hawthorne effect [45,47,85]. The novelty effect [86] applies to most of the included studies as they did not control for experience with new technologies or their affinity to them. Second, concerns about data access, storage or control, the ownership of AI-generated data, and its further use need to be clarified [87,88]. Standards should be derived and updated at appropriate intervals, especially new AI-generated knowledge based on employee’s personal information [89]. Transparency and high data protection regulation can increase adherence rates and reduce usage barriers [90]. In turn, we propose that future research should rely on reporting guidelines [76,91,92].

Of note, 1 strength of our review is the explanatory nature of the RQs and the systematic search strategy in this new field. Consequently, the heterogeneity of the identified studies might be considered a limitation. Different AI applications and technologies, the types of intervention, and the variety of workplace settings limit the conclusion significantly. Beyond this, the reporting of the types of AI-based algorithms and technologies used in the study are based on the authors’ self-reports. It is important to consider that the differentiation of the AI algorithm types cannot be made with a high degree of distinction.

Overall, this review underlines that AI in WHPP bears considerable potential but is not used fully at present. The results of our review offer a promising perspective on the predictive and personalized health paradigm shift in WHPP. Nevertheless, we conclude that current AI-related research in WHPP is still at the beginning, as it does not cover the scope of WHPP. The most salient research gaps can be found in lacking fields of WHPP and its subdomains, the predominantly ML-based algorithms and cross-sectional data, and the weak consideration of the work context. We believe we have contributed to future WHPP research by identifying these gaps and recommending future approaches. As AI applications are gaining an increasingly important role, we are convinced that future research will profit from an extended range of research in all fields of WHPP, longitudinal data, and the use of reporting guidelines.