Noncommunicable diseases (NCDs) are a major global health concern, responsible for more than 74% of deaths worldwide and significantly burdening health care systems [1]. In Europe, for example, cardiovascular diseases, cancer, respiratory diseases, and diabetes account for 75% of the disease burden within the European Union alone [2]. Beyond behavioral and environmental risk factors, growing evidence highlights the role of chronic psychological stress as a major contributor to the onset and progression of various NCDs [3-6] through mechanisms such as systemic inflammation, immune dysregulation, and metabolic dysfunction [6].

This burden is not confined to older populations but increasingly also affects younger vulnerable individuals, especially those with low socioeconomic status (SES)—a latent construct encompassing educational level, income, and occupational status [7]. Low-SES individuals are often disproportionately exposed to chronic stressors related to financial insecurity, precarious employment, and limited access to preventive health services [8-10]. Therefore, given this background, this work focuses on a digital stress reduction intervention as an upstream preventive measure for NCDs.

However, disparities in health care access also manifest in the “digital health divide” [8,11], where low-SES groups often lack access to preventive digital measures. While digital health interventions (DHIs) have been shown to positively influence proximal health outcomes such as patient engagement, self-awareness, and self-efficacy [12-15], they also contribute to improvements in distal health outcomes such as reduced stress [16,17], symptom reduction in depression or anxiety [18], or improved quality of life [19]. Moreover, DHIs offer several practical advantages such as overcoming geographical and temporal barriers [20], delivering engaging multimedia content [21-23], and enabling real-time data collection and biofeedback [24,25]. However, despite these benefits, socioeconomic disparities in their use [26-28] and evaluation persist [29]. Most notably, individuals with higher educational levels are more likely to use health apps than those with a lower educational level [11], meaning that underserved populations who stand to benefit the most from the scalable, personalized care that DHIs offer [29,30] are those least likely to use them.

Today, the delivery of DHIs is often skewed toward channels predominantly used by higher-SES individuals, such as workplace wellness initiatives and exclusive medical facilities, further hindering access [31]. Social media platforms, and influencers in particular, offer a promising means to reach vulnerable individuals either directly, for instance, via Facebook [32,33], or indirectly by upskilling (“influencing”) the communication of mental health content creators (“influencers”), such as on TikTok [34]. A growing body of observational studies in marketing, communication, and public health are further demonstrating that influencers can shape health-related attitudes and behaviors across a variety of contexts and populations [35-37]. Hence, working with influencers might constitute a promising strategy to promote “digital stealth health interventions” [38,39] and reach vulnerable individuals from lower social classes where they already are “healthifying their comfort zones” [40].

Beyond challenges in reaching vulnerable individuals, the design of a DHI itself plays a critical role in sustaining engagement, especially among vulnerable individuals. Low-SES individuals often face barriers such as limited digital literacy and health literacy or competing priorities, making it essential for the DHI to be engaging, relatable, and accessible, as well as relevant to their distinct needs, values, and resources [41].

Taken together, while DHIs hold immense potential, their successful implementation and impact depend on addressing the persistent challenges of reaching and engaging individuals of a low SES. Accordingly, our primary research questions (RQs) are as follows:

To effectively engage low-SES individuals with DHIs, it is important to recognize that vulnerable populations are not a homogeneous group but, rather, encompass a range of value orientations and social contexts each representing specific subpopulations. Indeed, while general economic aspects of SES remain a critical factor in understanding health disparities, recent sociological theories have shifted toward more nuanced frameworks that also account for individuals’ value orientations and lifestyle preferences. The SINUS Institute’s social milieu model, a widely used segmentation framework [42], captures these differences by categorizing populations based on both their social status (educational level, income, and occupational status) and basic values (cultural norms, aspirations, and worldviews) [43].

Figure 1 [43,44] illustrates how these social milieux can be mapped along two axes: (1) the vertical axis representing social status, ranging from low to high; and (2) the horizontal axis reflecting basic values, classified as traditional or conservative, modern or individualistic, and experimental or sensation oriented. These values are often shaped by the historical contexts in which individuals have lived and the experiences of different generations [42]. This model reveals that, even within lower-SES populations, there are diverse subgroups with varying lifestyle preferences and media consumption habits [44]. In this work, we focus on the 3 distinct social milieux characterizing populations of a lower social status: “traditionalists,” “consumer-materialists,” and “sensation-oriented individuals.”

Traditionalists are typically older, conservative, and oriented toward security and established social norms [43]. They are more inclined to engage with familiar and conventional media outlets such as television and classic social media [44].

Consumer-materialists find themselves in precarious living conditions, too, but aspire to the consumption standards of the middle class, seeking status and gratification through modern, individualistic experiences. They value modernization, individualism, self-actualization, and pleasure [43]. They are more likely to engage with social media, gaming, and video content as a means of personal fulfillment [44].

Sensation-oriented individuals are younger and driven by a desire for entertainment, novelty, and action. They gravitate toward immersive and fast-paced media such as computer games [44] and thrive on experiences that offer novelty and engagement, preferring highly visual and story-driven content [43].

Low-SES individuals tend to engage more with social media platforms [45], television (direct and streaming), and computer games than with traditional media outlets such as newspapers [44] (Figure 1 [43,44]). Accordingly, social media platforms offer a promising avenue for promoting DHIs and reaching vulnerable individuals. Content creators on social media in particular have the potential to serve as influencers, leveraging their relationships with their audience to encourage engagement with DHIs [34]. In 2023, major social media platforms such as YouTube, Instagram, and TikTok reported billions of daily active users [46] consuming vast amounts of content, making them ideal channels for health communication. Research further indicates that users on these platforms are strongly influenced by content creators when making purchasing or behavioral decisions [47,48], further supporting the potential of leveraging influencers to promote DHIs effectively.

However, leveraging influencers for health promotion does not come without challenges. Studies have highlighted that the impact of influencers can vary considerably depending on the credibility of the influencer, the topic, and the audience’s media habits and literacy, with some research showing limited or even counterproductive effects among certain groups [36,49,50]. Moreover, social media platforms have been criticized as fertile grounds for misinformation and health-related disinformation, with some influencers deliberately or inadvertently spreading inaccurate or misleading information to boost engagement or viewership (“clickbaiting”) [51,52].

Despite these limitations, influencer-based strategies remain underexplored in the context of embedding DHIs for vulnerable groups into their “comfort zones.” To our knowledge, there is no prior work that has assessed the effectiveness of influencer marketing when their content genre and audience values are systematically and conceptually aligned with the design of the DHI itself, particularly in the context of socioeconomically diverse social milieux.

We hypothesize that, by further matching the focal genre of an influencer’s content to the values of the targeted social milieux, DHIs can be positioned within the comfort zones of distinct social milieux. For example, influencers who produce real-life, lifestyle-oriented content may be more effective in promoting DHIs that are framed like conventional health apps to traditionalists, whereas influencers in the gaming community may be better suited for reaching sensation-oriented individuals to engage with DHIs packaged as a smartphone game. Taken together, we hypothesize that promoting a DHI through influencers whose content genre aligns with the basic values of a target social milieu will increase reach among vulnerable individuals (hypothesis 1).

Given our focus on DHIs in prevention health care settings and given the specific basic values (ie, conservative, modern or individual, and experimental) of the 3 target social milieux (ie, traditionalists, consumer-materialists, and sensation oriented individuals), we were eager to develop design guidelines (Figure 2) that would allow us to manipulate a DHI’s design along the spectrum of the social milieux’s basic values with regard to (1) the thematic framing of the app, (2) the social role embodied by the chatbot delivering the DHI, (3) the integration of gamification, and (4) storytelling and narrative elements—classic features in human-computer interaction and technology adoption research to enhance user engagement [53-55].

First, thematic framing involves embedding the intervention within a context that is similar to the target milieux’s values. The thematic coherence of a DHI can serve as a cognitive anchor, helping reduce distance and making it easier for users to relate to the intervention [56,57]. Such a thematic framing strategy might help a DHI feel less like a disruptive experience [58] and more like a natural extension of users’ daily media consumption habits [59]. We reasoned that, for instance, a DHI resembling a conventional health app might appeal to more conservative, security-oriented traditionalist individuals by evoking themes of familiar, predictable contents. In contrast, framing a DHI as a fitness influencer app focusing on self-expression may attract users who prioritize self-fulfillment. Similarly, a DHI designed like a game without any salient or obvious links to health might appeal to sensation-oriented individuals.

Second, social role theory suggests that the role played by a chatbot can influence user engagement [60]. When chatbots are designed to embody particular social roles (eg, health expert, coach, or relatable peer), they can activate social scripts and expectations aligned with the users’ value orientations [60]. In line with the thematic framing, traditional users may prefer a paternalistic expert that conveys authority and provides clear instructions (ie, a chatbot that offers step-by-step health advice, such as “Please follow the prescribed actions carefully to achieve the desired outcomes”), whereas individualistic users might respond better to a competitive coaching figure that motivates self-improvement (eg, “You’re on the right track! How about beating your last record?”). Sensation-seeking users, on the other hand, might engage more deeply with a story-based companion that integrates gamified elements (ie, “You’ve just unlocked a new level! Keep going to help the character complete the adventure!”).

Third, in line with the thematic framing, including gamification elements in DHIs for the consumer-materialists and sensation-oriented individuals has the potential to increase the experiential value for participants [61,62], potentially leading to higher commitment and adherence [53,63-65]. Gamification elements encompass quests and challenges or points, badges, and leaderboards in applications and for purposes beyond mere entertainment, such as education and health promotion [66].

Similarly, storytelling and narrative elements function as a mechanism for creating immersive and emotionally engaging experiences, potentially especially interesting for sensation-oriented low-SES individuals. Narratives have the power to draw users into an intervention by providing a framework for understanding complex concepts, building emotional connections, and enhancing the relatability of the content. Storytelling also allows for the creation of identifiable characters and scenarios, which can serve as points of reference for users to project their own experiences, thereby fostering a deeper personal connection to the intervention [54,55]. Each of these mechanisms—gamification, storytelling, and social roles—can be strategically used to optimize the thematic anchoring of a DHI and engagement by aligning the intervention’s design with the needs of specific low-SES social milieux. Taken together, we hypothesize that tailoring DHI design elements—specifically, thematic framing, social role, gamification, and storytelling—to reflect the value orientation of specific social milieux will enhance user engagement (hypothesis 2).

To answer our RQs, our study used a cross-sectional, quasi-experimental field study design to investigate the impact of differently designed versions of a health-promoting app that were promoted via influencers on the video streaming platform Twitch on audience reach and user engagement with the app.

To describe the study procedure, we need to, first, differentiate between three different parties involved in the study—(1) the study team, (2) the influencers, and (3) the end users targeted by our DHI (ie, our study participants=low-SES individuals)—and, second, divide the study into four parts—(1) development of the different app versions and study setup (refer to the Development of Study App Versions and Technical Implementation section), (2) recruitment and selection of the Twitch influencers (refer to the Recruitment and Selection of Twitch Streamers section), (3) reaching study participants through the streamers’ viewers (refer to the Recruitment of Study Participants via Influencers section), and (4) engagement with the study app by the target study participants (refer to the User Interactions With the Study App section).

Overall, we modeled 3 distinct versions of the DHI along the 3 low–social status social milieux’s guiding design principles outlined in Figure 1 [43,44] with regard to the following app elements: (1) graphical design, (2) the featured chatbot and its social role and conversational style, and (3) the graphical design of the guided slow-paced breathing training.

StressLess aimed to mimic state-of-the-art health-related apps such as Calm [68] or Headspace [69] and address the traditionalist milieu, who have conservative values and appreciate status quo–oriented apps; the focus of the app was primarily on the health and prevention aspect, and the design was moderate and plain. The breathing training was traditional without gamified components, which leads to rather low experiential or hedonic value.

Breeze aimed to address the value of pleasure and individualization of the consumer-materialist milieux. Breeze builds on the base of StressLess by maintaining a focus on the health and prevention aspect but imitates modern fitness and lifestyle influencers. The breathing training has gamified elements that increase the experiential value.

TragicKingdom was intended to address the hedonistic values of the sensation-oriented social milieu by implementing a storytelling approach with characters from the action role-playing game genre that leads the user through an adventure with the app character. The focus shifts from health and prevention to the storytelling aspect and the hedonistic and entertaining features of the app. The breathing training is also gamified and is integrated as a direct part of the story, which offers a high experiential value.

Figure 3 provides an overview of the differences among the 3 app versions. Details are outlined in the following sections.

The 3 DHI versions purposefully differed in terms of the overall look and feel of the app with regard to the name, logo, and color palette. While StressLess was toned down with calming and relaxing pastel colors; a rounded logo; and a conventional, reassuring name, Breeze was designed with bright, refreshing colors; a squared logo featuring the outline of a sailing boat as a reference to the storytelling features; and a modern, fun name. To highlight the storytelling and gamification features, TragicKingdom was designed using dark and mystical colors, a logo featuring a fantasy game–inspired emblem, and an immersive name. These themes were implemented thoroughly throughout all app and design components, including, for instance, the welcome screen, background screens, and chat message, but also in the design of the chatbot avatars and the breathing exercises. Figures 4A-4C illustratively show the similarities and differences in terms of design, the chatbots’ social roles and conversational style, and storyline based on a user interaction with the chatbots in the different app versions shortly before the breathing training starts. More screenshots of each app version are provided in Multimedia Appendix 1 (section A3).

The chatbots in the different app versions differed with regard to their avatar; the overarching storyline; the social role they represented; and their use of distinct interaction styles, wordings, and emojis. First, each version relied on a unique storyline—StressLess and Breeze featured a short introduction on the benefits of stress management (“patient education”), whereas TragicKingdom avoided any explicit educational contents and featured an adventure story instead. Second and third, the chatbots in the distinct app versions represented distinct social roles and used different interaction and linguistic styles accordingly. The chatbot character in StressLess was represented by a more distant “health expert” social role [60] that incorporated a paternalistic interaction style guiding the user through the intervention with a “top-down” approach, whereas the character in Breeze was framed as personal coach using a more deliberative interaction style characterized by joint decision-making approaches and the chatbot in TragicKingdom resembled a story character role. In all app versions, during onboarding, participants were allowed to choose between a feminine avatar called Alexandra or a male avatar called Alexander as previous chatbot studies have shown positive effects of letting individuals customize their chatbots [60]. Fourth, emojis were used by the chatbots in Breeze and TragicKingdom to emphasize text and express feelings. To further amplify the story in TragicKingdom, some photographs were also shared via the chat (eg, a picture of the character who needed to be rescued by the user and the map of the island).

Preventing NCDs involves primary, secondary, and tertiary measures, with this study focusing on primary prevention and educating vulnerable individuals on maintaining a healthy lifestyle [70,71]. Psychological stress is a significant factor in the development of many NCDs such as hypertension, diabetes, and cardiovascular disease [3,4,72], triggering harmful mechanisms such as the fight-or-flight response when prolonged [5,73,74]. Research shows that stress can alter biological processes relevant to diseases, impacting health status [3].

To mitigate stress, slow-paced breathing training has proven effective in improving psychological and physiological well-being [75]; reducing the risks of mental and chronic physical health conditions, including depression [76], anxiety [77], stress [78], chronic obstructive pulmonary disease [79], type 2 diabetes mellitus [80], chronic pain [81], and hypertension [82]; and managing acute and chronic stress [75,83]. Digital, biofeedback-enhanced slow-paced breathing training using biosignals to guide users through the exercise has shown particular promise in stress reduction and user engagement, especially for less motivated or vulnerable individuals [67]. However, designing effective biofeedback breathing training with gamified components remains challenging [64].

The slow-paced breathing training was adapted in terms of color, functionalities, and design in line with the overarching storylines and design themes of the app versions. The training in StressLess mimicked state-of-the art meditation and breathing exercise apps through a circle that expanded and shrank with the inhalation and exhalation with no further gamified components or biofeedback (Figure 5A). Moreover, the design was kept simple in a calming blue color scheme.

Alternatively, Breeze and TragicKingdom included a gamified and lavishly designed slow-paced breathing training with real-time biofeedback (Figures 5B and 5C). The focal design element of the training was a sailing boat that accelerated through the breathing pattern of the user and reached new landscapes; this element provided a gamified component and increased motivation to continue. On the basis of previous research, the landscapes included gardens, trees, and green spaces to set a natural environment and foster a calming and meditative effect on the user. The game used the sailing boat and the wind as dynamic objects within the static environment. During the inhalation phase, the wind moved against the direction of the boat, and during the exhalation phase, the wind moved toward the boat to inflate the sails. These movements illustrated the breathing patterns and guided the user through the breathing cycles. The wind strength symbolizes biofeedback, and therefore, the quality of the breaths was illustrated by the movement of the sail, the boat’s speed, the size of the waves, and the increasing number and speed of particles [64,67,84]. Previous studies have proven the functionality and applicability of the aforementioned breathing training. For example, gamified training has a higher perceived enjoyment from participants than a standard breathing exercise, such as using a circle (StressLess). Studies have also reported that gamified biofeedback increases the experimental values and fosters engagement [64,67]. These design distinctions led to significant differences in terms of experiential value, design, and biofeedback.

The agency reported that they contacted approximately 80 to 90 streamers before recruiting the final 8 employed in the study. The streamers’ follower numbers ranged from 360 to 44,225 (based on archival TwitchTracker data from February 20, 2022). However, the number of followers does not necessarily directly reflect a streamer’s average live viewer numbers and the engagement of the community [106].

The included Twitch streamers advertised the study between February 2022 and April 2022. The period was increased from the original plan due to delays and cancellations from the streamers.

The average viewers during an advertisement stream ranged from 16 to 3700 (mean 506.7, SD 1134.3). In total, the 8 streamers performed 25 advertisement streams. Overall, we recorded 273 downloads; 172 (63%) downloads could be allocated to a streamer ID based on information recorded via the deep link data, another 30 (11%) downloads could be allocated to a streamer based on manual requests, and 18 (6.6%) could be allocated to a streamer due to timing, resulting in a final sample of 220 total downloads. A total of 9.9% (27/273) of the downloads could not be considered as they did not include the streamer ID or could not be attributed to a streamer based on timing. Another 9.5% (26/273) of the downloads were from the participating streamer themselves or were made before the streamer’s first advertisement stream and, therefore, excluded.

Overall, the streamers performed the advertisement well and in accordance with the requirements discussed in the Recruitment of Study Participants via Influencers section. All streamers directly advertised the apps at least once during a stream; showed the promotion video; displayed the banner; and showed their smartphone with the app installed, including occasionally screen sharing the app live during the stream to enable the viewers to directly experience the intervention. All streamers streamed in the categories IRL or games. In general, the interest from the streamers and the audience seemed high.

However, it should also be noted that the precise reach of each streamer could not be fully determined due to idiosyncrasies in how individual streamers further shared their content outside of the stream (eg, uploading the stream to be rewatched on YouTube and publishing static advertisements on other social media platforms). No data were available on the reach of these marketing actions. Furthermore, streamer ID 4, for instance, initially advertised the intervention only for iOS devices (but later clarified that the intervention worked for all operating systems), streamer ID 5 executed most of the advertisements early in the morning (eg, 5 AM), and streamers ID 1 and 5 demonstrated the interventions live by screen sharing them during the stream. As expected, the advertisements by each streamer furthermore differed slightly in timing, length, style, and wording. The collaboration with the agency as well as with the streamers went well. However, coordinating 8 different streamers entailed a high communication load, short-notice changes, and adaptations. Impressions of and notes on the advertisement streams of each streamer are included in Multimedia Appendix 1 (section C3).

Table 1 shows an overview of the number of streams; average viewers per stream; number of generated downloads; and average reach per stream, streamer, genre, and app version. In total, 220 downloads were counted, of which 92 (41.8%) were for StressLess, 66 (30%) were for Breeze, and 62 (28.2%) were for TragicKingdom.

Average reach per stream amounted to 16.2%, ranging from 0.1% (2/3526; ie, stream ID 4-3 and 5/3700, for stream ID 4-4) to 56% (9/16; stream ID 6-3). A Kruskal-Wallis test found significant differences between app conditions (ꭓ²₂=8.0; P=.02), with average reach ratios of 8.1% for streams promoting StressLess, 14% for Breeze, and 28.4% for TragicKingdom (Table 1). Pairwise comparisons revealed significant differences between StressLess and TragicKingdom (P=.005) but not between Breeze and TragicKingdom (P=.15) or between StressLess and Breeze (P=.23). These findings indicate that TragicKingdom outperformed StressLess and Breeze in terms of reach per stream, whereas there was no significant difference in performance between StressLess and Breeze.

The results further highlight a match or mismatch pattern between the dominant genres of the influencers’ content and the app version they promoted. The “match” condition referred to when the content genre of the influencer aligned with the thematic framing of the DHI. The match between the games genre and the TragicKingdom intervention resulted in the highest reach ratio of 28.5% (SD 17.6%). In contrast, StressLess, which represented a mismatch in this genre, achieved a significantly lower reach ratio of 14.9% (SD 9.3%). On the other hand, in the IRL genre, the reach ratio for Breeze, which was a match for this genre, amounted to 14% (SD 10.5%), whereas StressLess, a mismatch, resulted in only 8.1% (SD 9.3%) reach. This pattern suggests that aligning the genre of the influencer with the thematic design and framing of the intervention can enhance user engagement, confirming hypothesis 1.

Overall, the results presented in this paper show that using Twitch for digital health recruitment is highly promising, moving far beyond some previously published reach ratios for DHIs using traditional social media websites (eg, 0.02% [108] and 0.1% [109]). However, balanced against this potential are occasionally low reach ratios by certain streamers (eg, 2/3526, 0.1%; stream 4-3 of streamer ID 4), which, by comparison, are no better than direct email campaigns to known contacts (eg, 2.5% using email with health insurance customers [110]), which are far simpler and cheaper to use. Therefore, the results presented in this paper indicate a mixed picture—while influencer-based recruitment strategies on Twitch can reach large audiences and may be a highly effective digital health recruitment tool, simply adopting an influencer strategy is not enough to guarantee success.

However, an important caveat in interpreting these findings were varying levels of engagement with the actual intervention content following app download. While initially high levels of engagement were found (ie, downloading the app and beginning the interaction with the chatbot), a substantial drop-off was evidenced, particularly during the gamified breathing training, where only 7.7% (17/220) of the participants completed the training in full. Adherence is a known challenge in digital health [111], and our findings show that, despite the potential of streaming platforms such as Twitch to recruit participants, it does not provide a panacea to resolve the well-documented dropout rates established across digital health domains [112]. While incentives—such as giveaways or monetary compensation—have been shown to increase adherence in digital health studies [95], their scalability and ethical applicability remain limited in public health domains, particularly in the context of prevention, in which no established reimbursement structures exist. For instance, Germany’s digital health application framework only covers therapeutic interventions for individuals with a formal diagnosis [113-115], underscoring the absence of a structured market for preventive DHIs. Nonetheless, incentives embedded within influencer-driven formats could represent a promising direction for future research.

Reasoning as to why these dropouts occurred, we offer several explanations. First, our rationale for using the Twitch platform was the likely presence of 2 risk factors among its audience, namely, habitual video game [116] and social media use [117]. However, the apps presented in our study may have been less appealing than the professionally developed video games or social media activities present on the platform, subsequently leading participants to return to Twitch rather than complete the breathing exercise in full. Second, relatedly, there is a strong tendency toward immediate gratification when using digital media [118], particularly among the younger demographics [119]. Thus, while the health benefits of the apps were presented by the streamers, these benefits may have been too intangible, contributing to dropouts. Third, the timing of DHIs is key [120]. However, participants were using Twitch, an entertainment platform used for recreation and relaxation purposes [121]. Thus, while participants may have indeed seen the value in the stress-based biofeedback game (leading to their download), they might not have been acutely stressed in the first place and, thus, had no reason to complete the breathing exercise at that time. We cannot rule out the influence of technical issues during the breathing training as no crash diagnostics or in-app analytics were available (which should be included in future studies) to distinguish disengagement from technical dropouts.

As a final point, while the recruitment strategy effectively reached a large audience (N=220) with an overall average reach ratio of 16.2%, we were unable to confirm the hypothesis that Twitch could engage low-SES individuals within their comfort zones due to a lack of survey responses. In contrast to other social media platforms, Twitch does not allow for the capturing of detailed audience metrics, which is perhaps why most extant research has focused on influencer-level analyses [122-124] rather than on the audience. In any case, because of this, we cannot assess the extent to which vulnerable individuals were reached despite our successes in overall recruitment. In hindsight, and for future research, we recommend integrating key SES-related screening questions early in the intervention flow (eg, during onboarding) or, alternatively, implementing a separate pre- or postintervention survey dedicated to SES assessment. Such an approach would allow for more robust verification of whether the intended vulnerable populations are being reached while minimizing disruption to the intervention experience itself.

On the basis of the findings presented in this paper, alongside those presented in extant research, we propose 3 factors governing the effectiveness of adopting an influencer-based strategy using streamers on Twitch for recruitment and health promotion purposes.

First, influencers vary greatly in terms of their communicative, interpersonal style [125], and this individual-level heterogeneity can lead to both differences in the ability to promote health content effectively and likely differences in the makeup of the audience. While we were unable to capture the audience background in this domain (due to constraints of Twitch and low survey response rates; also refer to the Limitations section), communication style has been found to be an important determinant in the recruitment for novel digital health tools [126], and matching health promotion content with the expectations of vulnerable individuals has a long history in health care research [127]. Indeed, in nonhealth domains, the communicative style of influencers has been shown to affect both advertising relevance and audience expectations [128]. Thus, we reason that carefully selecting influencers whose public personas and communication styles align well with the intervention’s goals and the values of the targeted social milieu is likely to lead to higher effectiveness. In addition, upskilling, educating, or “influencing influencers,” as described in the study by Motta et al [34], may further improve their ability to convey health-related content in a way that is both authentic and motivating.

Second, the effectiveness of Twitch influencers was found to be more pronounced among smaller-scale streamers or microinfluencers. These influencers, who typically have fewer followers, appear to foster a more authentic connection via the greater time they can afford in personalized interactions with followers [106]. Indeed, greater interactivity is a common factor in the success of various novel technological platforms deployed in digital health [129], such as self-service technologies [130] or chatbots [131], and a strength of social media–based marketing more generally [132]. Thus, we reason that this rationale similarly extends to Twitch or influencer marketing, where greater attention spent by an influencer on each participant facilitated “parasocial relationship” (ie, an imagined intimate relationship with the streamer) building, a phenomenon documented on Twitch previously [121]. Accordingly, future interventions could explore the integration of real-time interaction between influencers and participants (eg, live questions and answers or feedback loops during the intervention) to further strengthen social presence and motivation.

Third, the streamers mostly streamed content in the categories games and IRL and for varying durations. Therefore, it is reasonable to assume that both the category selected and the duration for which the gamified breathing exercise was presented influenced downloads. Indeed, data gathered showed that longer streams were actually counterproductive in generating downloads, likely due to the “try before you buy” phenomenon on Twitch (ie, users had vicariously played the entire app content and, thus, had no further need to download the app) [121]. Similarly, we reason that the choice of categories (eg, game) or specific subcategories (eg, comparing Minecraft to Grand Theft Auto) that differ in terms of congruency to a relaxing breathing exercise may have influenced the downloads.

This study featured real streamers promoting real health apps, capturing highly relevant behavioral data (ie, app downloads and app engagement) and, thus, possesses a high degree of external validity for digital health recruitment. That being said, there are also a number of limitations in study design and setup that should be acknowledged when interpreting the results.

First, as noted, due to limited survey responses after the intervention, we were unable to assess the audience of those downloading the app and, thus, could not assess demographics or the vulnerability status of the users, a core aim of this research project. Moreover, while the DHI designs were tailored to align with the value orientations of the SINUS Institute social milieux, they were not directly cocreated with the influencers’ audiences or informed by specific audience needs assessments. However, this choice reflects our study’s focus on testing theoretical reach and engagement mechanisms in a controlled field setting using 2 previously evaluated versions (StressLess and Breeze [64,67,133]) and 1 newly developed, dedicated version (TragicKingdom) designed with the Twitch platform’s audience profile in mind (also refer to the Future Research section).

Second, due to the inherent nature of live streaming, health messages were not delivered in a fully standardized manner by streamers, which likely introduced additional variability in download frequencies. While we compensated this by providing standardized scripts and content to streamers, streamers were also encouraged to retain their personal communicative style so that we could organically assess the potential of streamers to promote health-related content effectively. In addition, while we used 2 streamers per genre to balance out strong individual effects unique to a single streamer, it is possible that the audiences varied between each of these streamers in terms of their health vulnerability [124], in turn influencing downloads. Together, these issues critically highlight the inherent challenges of aligning digital channels and DHI designs with the value orientations of specific social milieux and the need to consider distinct platforms for distinct milieux (also refer to the Future Research section).

Third, the 3 experimental conditions varied several elements simultaneously (eg, related to the chatbot, video style, and influencer) between conditions. Although this was necessary to match DHI design with the social milieux established by the SINUS Institute [43], this research design meant that it was not possible to fully disentangle the relative effects of each experimental condition individually. Relatedly, our design of the “sensation-oriented” TragicKingdom version raises important considerations about what constitutes a DHI in the first place. While this version did not explicitly frame the activity as health related, it was designed to promote health-relevant behavior through gamified engagement with stress-reducing breathing exercises. While this broadly aligns with emerging research on so-called stealth health interventions [38,39], which emphasizes the potential of implicit, play-based strategies, future research should further explore the effectiveness and acceptance of such approaches in comparison to more explicitly health-framed interventions.

Finally, while the interventions were designed as low burden, intuitive, and accessible with simplified language and scripted chatbot dialogues tailored for laypersons, we acknowledge that digital health literacy remains a key factor influencing access, understanding, and sustained engagement, particularly among socioeconomically vulnerable groups. Given the well-documented disparities in digital health literacy [27], future research might explore combining milieu-specific DHI tailoring and influencer-based outreach with explicit DHI support measures such as digital onboarding or user education strategies to further reduce barriers and ensure equitable uptake.

Understanding the recruitment for DHIs is an emerging research area that has faced increasing scrutiny in recent years. Nevertheless, the topic remains relatively immature when compared to the wealth of data on DHIs and treatment outcomes. Therefore, we offer several research directions that should prove useful in expanding this domain.

First, investigating congruency between the promoted app or intervention and the personality of the influencer would likely prove fruitful. Research has shown that congruency between social media advertisement content and audience values or demographics has a strong influence in their effectiveness as a recruitment tool [134-136]. Thus, understanding how certain streamers may more effectively promote health-related content among certain audiences (particularly those exhibiting higher vulnerability) would enable more effective targeting. In this vein, future research could build on our findings by incorporating further formative research, audience profiling, and participatory design methods to ensure that the interventions even more closely reflect the lived experiences, preferences, and perceived needs of the target groups. In a similar vein, qualitative postintervention assessments or follow-up interviews could provide valuable insights into user experiences and engagement trajectories, especially to better understand dropout reasons.

Second, a comparison of different social media platforms for recruiting vulnerable participants would be highly welcome. While reviews are available on recruitment and health promotion efforts across various social media platforms [137], these are hindered by being made up of papers that vary widely in scope (eg, health domain, audiences, funding, duration, and other factors). Therefore, it would be useful to have a single study that controls for these various factors and uses multiple social media platforms simultaneously to recruit participants. This would enable a relative ranking of the strengths or weaknesses of such platforms among different audiences.

Third, the use of virtual influencers (ie, artificially created characters combining the use of computer-generated images and artificial intelligence [138]) for health promotion purposes would be an interesting new research direction. Virtual influencers have found success in commercial settings for advertising products on social media [139] and were even experimentally used by the World Health Organization during the COVID-19 pandemic [140]. Therefore, we reason that further exploration of their capabilities for health promotion among vulnerable groups may open up a new method of recruitment. This would be particularly interesting considering the rise of increasingly powerful large language models that, once integrated with virtual influencers, will inevitably improve their realism, authenticity, and ability to promote health content [141].

Finally, to the authors’ knowledge, there is not yet a large-scale study that integrates social media recruitment, intervention adherence, and longitudinal health outcomes. In the context of this study, this could have been achieved by extending the study design and capturing relevant biomarkers gathered during the biofeedback breathing training. Doing so would greatly improve our understanding of how social media can be leveraged to actually improve health outcomes over the long term and would be particularly impactful if targeted toward vulnerable individuals.

In this study, we successfully demonstrated the potential of Twitch for digital health recruitment, with high reach ratios that, at times, vastly outperformed those of other digital health recruitment tools. Nonetheless, we were unable to verify the assumption that Twitch could target and engage vulnerable individuals within their comfort zones due to a high number of dropouts. Moving forward, we encourage future research to gather both more audience-level and more streamer-level data so that innovative recruitment strategies can be developed to meet vulnerable individuals within their comfort zones. Doing so will ensure the equitable rollout of DHIs across the population; this study takes an important first step toward this goal.