We have conducted a scoping review to capture current evidence on HFs in AI-driven digital solutions for increasing PA. A scoping review is a systematic approach used to map and synthesize existing literature on a broad topic, providing an overview of key concepts, sources, and knowledge gaps. Our review followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [21].

We have searched for publications including keywords related to PA (ie, “physical activity;” “exercise;” “active lifestyle;” “sedentary behaviour;” “inactivity;” “resistance training;” “exergaming;” “walking;” “swimming;” “jogging;” “climbing”), artificial intelligence (ie, “artificial intelligence;” “AI;” “machine learning;” “deep learning;” “natural language processing;” “neural networks;” “sentiment analysis”), and human factors (ie, “usability;” “task performance;” “satisfaction;” “workload;” “human errors;” “user perception;” “cognitive factors;” “mental model;” “context awareness;” “automation bias;” “teamworking;” “user experience;” “acceptance;” “acceptability;” “task analysis;” “handover;” “patient interaction;” “human factors;” “ergonomics”) in their titles and abstracts. No language or year limitations were used. The full search strategy can be found in Multimedia Appendix 1.

The data search was performed on August 29, 2023. The database search was done by a single author (EG) and covered PubMed, Embase, and IEEE Xplore. Another author (DL) carried out a search on Google Scholar and selected the first 100 entries. Finally, DL used a snowballing approach to identify additional relevant studies cited in only the included publications.

Inclusion and exclusion criteria are presented in Textbox 1.

All references were uploaded to EndNote (version 20.6; Clarivate) [22] and Rayyan (Qatar Computing Research Institute) [23]. After duplicates were removed, 2 authors (EG and DL) independently assessed the eligibility of the remaining publications by checking their titles and abstracts. Two additional authors (KD and OR-R) checked the full text of the eligible papers after the title and abstract screening. After the full-text screening, the selected papers were included in a qualitative synthesis.

Two authors (KD and OR-R) extracted the following data: bibliographic information (publication year and country); study characteristics (study design, type of evaluation, research methods, primary and secondary measures, materials, and theoretical foundations); population (number of participants, age, and gender); intervention (intervention design, duration, and follow-ups); comparison (control group or groups and pre-post evaluation or other); outcomes (primary and secondary outcomes); and AI-related information (technology type, main purpose, platform, and HFs).

OR-R identified and assigned codes representative of the main purpose of the AI model implemented in each of the systems studied. The 3 main purposes of the AI models implemented in the studied systems were identified as personalization, communication, and human activity recognition. Personalization includes all AI models analyzed whose main purpose was to adapt the digital solution or intervention to the patient’s needs, conditions, and preferences. The second group includes models that enabled a communication pathway with patients. Finally, human activity recognition includes all models that enable the detection of user behaviors, particularly PA. OR-R and KD reviewed the assigned codes and created a classification of these by consensus.

The certainty of the evidence on the outcomes was assessed by a single author (EG) by drawing on the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) criteria [24] and verified by the rest of the coauthors.

A total of 2076 articles were identified in the data search. After removing duplicates, 1979 titles and abstracts were screened for eligibility. Of those, 13 publications met the inclusion criteria [25-37]. The snowballing approach identified 2 additional publications [38,39]; therefore, the final number of publications included in this review was 15 (Figure 1 shows the PRISMA [Preferred Reporting Items for Systematic Reviews and Meta-Analyses] flowchart).

The list of publications excluded during the full-text search and the reasons for their exclusion are reported in Multimedia Appendix 2.

The 15 included articles were published between 2015 and 2023. Countries of origin of these studies were: United States (n=3) [25,26,36], Australia (n=2) [32,38], South Korea (n=2) [29,35], the Netherlands (n=1) [31], Italy (n=1) [27], Belgium and Italy (n=1) [30], Thailand (n=1) [37], and Taiwan (n=1) [33]. A total of 3 publications did not specify in which country the study was performed [28,34,39].

Regarding the study design, 8 publications followed a quasi-experimental approach [26-29,32,35,38,39], 5 were randomized controlled trials [25,30,33,36,37], and 2 were exploratory studies [31,34].

Only 4 of the 15 included publications explicitly mentioned their theoretical foundations. The following theoretical approaches were cited in these 4 publications: the Fogg Model for behavior change [25,31], Capability, Opportunity, and Motivation model of Behavior [32]; learning theory [25], social cognitive theory [25], and the Transtheoretical Model [39].

The main technical characteristics of the 15 included publications are presented in Table 1.

The most common AI technology type was recommender systems, described in 10 of the 15 included publications [25,26,29-31,33,35-37,39]. In addition to the recommender system, one of these publications also included computer vision [29]. Conversational agents were the second most used AI technology, as described in 4 publications [27,28,32,38]. One of them was integrated into a social media platform, namely Slack [28]. One study tested human activity recognition [34]. Details of the AI technology, systems, or platforms used in the included studies are summarized in Table 1.

Regarding the considered HFs, the most commonly evaluated were acceptability [26,32,34,38,39] and satisfaction [29,34-37], both reported in 5 publications. Usability was the next most considered and evaluated HF, as reported in 4 papers [29,31-33]. Usefulness was assessed in 2 publications [26,29]. Other considered and evaluated HFs were engagement [38], UX or individual perception [25], and task performance [34].

The studies that resulted in increased PA and had a moderate certainty of evidence were chatbot systems with integrated recommender systems. Although the usability of some of those systems was considered poor [32], they were perceived positively. Several papers gained interesting results regarding HFs. Given the variety of systems, a generalization for all 15 studies is difficult.

The automated collection of data needed for personalization and recommendation is an important aspect. In total, 5 systems involved fitness trackers [26,30,32,35,38] to enable automated data collection. They can be grouped into mobile-based activity tracking using movement sensors in the phone [25], dedicated fitness trackers [30,32,38], specifically an accelerometer in the wristband [30], Fitbit Flex 1 activity tracker (Fitbit LLC) [32], and Garmin Vivofit4 tracker (Garmin) [38], and smartwatches [35]. Rabbi et al [25] concluded that automated data collection would be useful. The studies involving chatbots concluded that users have high expectations regarding the chatbot’s knowledge and capabilities [28]. Human likeness is reported as a success factor of such systems. Relevant aspects leading to the efficacy of the system include the human-like qualities of the chatbot and the personalization of the suggestions [28,32,39], that is, chatbots or digital assistants should have a personality, have humor, be able to act with spontaneous behavior, and in a diverse, nonrepetitive manner [28,32]. They should provide the correct answers. For successful recommendations, it is essential to learn the personal preferences of users so that suggestions can be made that fit into personal routines and lifestyles [39].

Even a combination of human agents and digital agents was reported to be better accepted than pure virtual support [27]. Beyond that, access to a system anywhere and anytime is well perceived [31]—and this is reflected by the fact that most systems included in this study are delivered as mobile apps (instead of desktop apps). Exercises and recommendations are successful in this setting when they can be easily integrated into the daily lives of the users [31].

A total of 899 individuals participated in the included publications. Of those, 60.7% (546) were female participants. The reported average ages of these participants ranged from 18.7 to 84.4 years. In total, 6 out of the 15 studies tested their solutions on participants with mean ages of around 50 years or older [28,31-33,37,38], while 6 studies predominantly included participants with a mean age of 40 years or younger [25,27,29,34,36,39]. Two studies did not specify the gender or age of participants [30,35].

The intervention of the included studies lasted between 2 and 26 weeks.

Prepost evaluations were carried out in 6 of the publications to evaluate the impact of the AI-driven intervention [26,29,32,35,38,39]. In 4 publications, control groups were used to assess the impact [25,30,34,36]. In 5 of the publications, the comparison methods used to assess the impact of the AI-driven intervention on increasing PA were not clearly reported [27,28,31,33,37].

The effectiveness of AI-driven technologies for increasing PA was shown in 5 publications [28,32,36,38,39]. Three of these publications tested conversational agents [28,32,38], while the other 2 focused on recommender systems [36,39]. The analysis, based on GRADE guidelines, found moderate certainty in the evidence supporting this statement. Further details about the proven effect of these studies and the certainty of the evidence on these findings are reported in Table 2.

In total, 4 of the included articles also showed that AI technologies have an effect on changing PA behavior (ie, feeling more stimulated to engage in PAs, change in walking behavior, improved behavior related to PA, or improved ability to perform PA) [25,30,35,37]. All these publications were recommender systems [25,30,35,37] and found a positive effect of AI-driven technology on changing PA behavior. However, the analysis, based on GRADE guidelines, found very low certainty evidence supporting this statement.

In this scoping review, we aimed to identify and describe the current evidence on HFs in AI-driven digital solutions for increasing PA. The results showed that the most common AI technology used in digital solutions for PA was recommender systems, followed by conversational agents. User acceptability and satisfaction were the most commonly evaluated HFs in the included studies. Some studies also evaluated the usability of AI-driven digital solutions for PA.

We have identified studies that provide evidence that AI-driven digital technologies have the potential to increase PA (eg, number of steps, distance walked, or time spent on PA). Furthermore, AI-driven technology, particularly recommender systems and chatbots, seems to have the potential to influence changes in PA behavior. Although these studies offer valuable insights by demonstrating positive outcomes through various AI-driven technologies for enhancing PA, the evidence is still very limited. The main findings are presented in Table 3.

In the included studies, we recognized several benefits of AI integrated into digital solutions for increasing PA, such as the ability to adapt the solution to the patient’s physical capacity, current activity, and psychological profile [8,11,30]. AI can monitor activity and inactivity and predict bodily occurrences, which is especially relevant for older people [40]. AI can also simulate the role of a personal trainer, provide guidance, form correction, and motivation [37] through voice- or text-based interactions. Users can receive real-time feedback and support during their workouts [8,10] which would be difficult to achieve with non-AI digital solutions. AI algorithms can analyze user data such as fitness levels, health conditions, and preferences and provide personalized exercise recommendations [11]. The activities or other suggestions are tailored to the specific needs and goals of the user, increasing the likelihood of adherence. Real-time feedback can be shared with the user. Previous studies found that activity tracking combined with real-time, personalized text messages can significantly increase PA and further affirm text messaging as an effective health behavior modifier [10-12]. However, in our review, researchers concluded that their solution did not achieve sufficient adherence to the exercise program [28,30]. The entire potential of personalization techniques has not yet been implemented in the solutions, as Luštrek et al [30] concluded that personalization, simplicity, ease of use, and avoiding information overload could be improved.

AI algorithms can continuously learn from user interactions and feedback to refine and improve the UX. This iterative process leads to more effective and engaging solutions over time. For example, the continuous interaction that chatbots can provide was reported to be useful in helping users increase regular PA and in helping them stay motivated to participate in PA [32]. Studies have already found that the human-likeness or anthropomorphisms of a chatbot increase the likelihood that users comply with the chatbot’s recommendations [18]. Roy and Naidoo [17] found that human qualities like warmth and competence are contributing to a positive UX and possibly to an increased adherence to the digital solution [17].

We only found 5 studies involving sensors to measure PA [26,30,32,35,38]. Dedicated fitness trackers seem to be more prominent to be involved in solutions increasing PA. Mobile-based activity tracking and smartwatches were only implemented in one solution. A reason might be that users prefer to use systems they already use; that is, integration with existing tools like fitness trackers is desired by users, as found by Wang et al [41]. The landscape of wearables and sensors that could be used for PA tracking is much larger than was found in our research [42]. The integration of sensors with AI could help analyze the data streams and promote an increase in PA [43]. Additionally, it could assist in monitoring PA among individuals affected by health conditions [44-46]. We hypothesize that existing research focuses on sensors that are well-known, not very intrusive, and therefore probably more accepted by users of solutions for increasing PA.

AI can gamify the fitness experience by setting challenges, goals, and rewards. Users are motivated to increase PA by earning points, competing with friends, or unlocking achievements. Xu et al [16] found in their review that gamification interventions could increase PA participation. Interestingly, none of our included studies explicitly reported about gamification elements.

In total, 5 of the included studies provide moderate evidence of AI’s effectiveness for increasing PA [28,32,36,38,39]. However, these studies involve short interventions lasting from 6 to 12 weeks. Hence, the significant effect might be influenced by this brief follow-up period, similar to other mHealth interventions [13]. The estimated time needed to form habits of complex behaviors such as exercise behavior is 12 weeks [47]. Thus, longer intervention studies are needed to assess the potential long-term effectiveness of AI-driven technologies for increasing PA.

Out of the 260 participants in these 5 studies [28,32,36,38,39], 72.3% (188) of them were women, and the majority were aged between 40 and 50 years. Further studies are needed to investigate the effects of these AI-driven technologies on participants with different sociodemographic characteristics, as well as those with health conditions for which exercise aids in managing the disease and preventing complications [1,2,4].

There is very limited and low-quality evidence supporting the impact of AI-driven technologies on changing PA behavior and the ability to perform such behavior [25,30,35,37]. In these cases, the durations of the interventions varied, ranging from as short as 3 to 4 weeks [25,37], to as long as 20 to 26 weeks [30,35]. Similar to previous cases, the majority of participants in these 4 studies were women, comprising 83.7% (82/98) reported participants. While research indicates that gender is one of the factors influencing the use of health-related technologies [48,49], technologies aimed at increasing PA should be tested, personalized, and accessible for all demographic groups.

Most of the included AI-based PA systems showed positive results in terms of HFs related to their use. However, no study aimed to evaluate how the AI component could independently influence HFs such as user acceptance, perceived ease of use, or perceived usefulness. Many studies used AI techniques to personalize the PA system based on the authors’ assumptions about the persuasive power of personalization that could lead to greater motivation and thus result in greater intention to use, adoption, and engagement. However, no study has tested these hypotheses. In this regard, more research is still needed to identify the role of AI components in HFs affecting PA systems. In addition, no study has focused on whether the inclusion of AI could lead to a change in the role of HFs, as has been the case with traditional technologies.

There were some identified limitations in this scoping review. Even though we did not have a language limitation in the search strategy, all the included studies were in English. Therefore, we could have missed relevant AI-driven solutions published in other languages. The included studies were mainly from diverse high-income countries, restricting generalization to low- and middle-income countries. In addition, the studies included in the scoping review had an intervention period of a maximum of 26 weeks, showing only the short-term effect of the AI-driven solutions. All studies were included in the review, irrespective of the assessed quality of the evidence. However, the results of the included studies were reported separately according to the quality of the evidence, minimizing misinterpretation of the data.

This study synthesized current evidence on the effectiveness and potential of AI-driven digital solutions for increasing PA. Although the included studies offer valuable insights by demonstrating positive outcomes through various AI-driven technologies for enhancing PA, the evidence is still very limited. While some studies demonstrated moderate evidence of AI’s effectiveness in increasing PA, these interventions were typically short-term. Longer-term studies are necessary to assess the sustained impact of AI-driven technologies on behavior change and habit formation. Additionally, further research is needed to investigate the effects of AI-driven interventions on diverse populations, including individuals with varying sociodemographic characteristics and conditions. Moreover, the evidence regarding the impact of AI-driven technologies on changing PA behavior remains limited and of low quality. There is a need for rigorous studies to evaluate the effectiveness of these interventions, particularly in terms of their ability to induce long-term behavior change. Furthermore, while most AI-based PA systems demonstrated positive results in terms of UX, there is a lack of research focusing on the independent influence of AI components on HFs, such as user acceptance and perceived usefulness. Additionally, more investigation is required to understand how the inclusion of AI may alter the role of HFs in PA systems compared to traditional technologies.

In conclusion, while AI-driven digital solutions hold significant promise for promoting PA and improving public health outcomes, addressing these limitations and challenges will be crucial for maximizing their effectiveness and accessibility. Continued research efforts in these areas are essential for advancing our understanding of the role of AI in PA promotion and ensuring the development of evidence-based interventions that benefit diverse populations.