The current systematic review thematically synthesized relevant published, peer-reviewed studies to identify audience-centered approaches applied in VHA-based intervention development. We selected thematic synthesis as our synthesis method because it allows for integrating findings across different study types (eg, qualitative, quantitative, and mixed methods studies) [30,31]. The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines provide international standards for reporting systematic reviews. This review followed the PRISMA 2020 checklist for systematic reviews [32] for its methodology. The following methodological steps were taken to identify, screen, assess, and synthesize relevant literature for this review.

Studies were included if they were primary research, published in a peer-reviewed journal, and written or available in English. All studies were required to contain research regarding the development or evaluation or both of VHA to promote cancer prevention or screening or both among adult populations in a primary care setting. We also included those studies that focused on the audience-facing aspects of the intervention, for example, VHA-audience interaction, message content, look, and feel of the intervention.

Studies were excluded if their sample focused exclusively on cancer treatment and survivorship or on the technological development of the VHA (ie, algorithms, programming language, and software development). Studies were also excluded if they focused on primary prevention mechanisms, such as smoking cessation without explicitly mentioning cancer prevention as an outcome. In addition, studies were excluded if the full texts were not available in English, the studies were literature reviews themselves, were not published in peer-reviewed academic journals, or were editorials or commentaries.

We searched 4 electronic databases between January 12, 2022, and January 14, 2022: MEDLINE (PubMed), Academic Search Premier (EBSCOhost), Embase (Embase), and Web of Science (Web of Science).

The search strategy was developed collaboratively by the authors (including 1 systematic review expert, CLB and 1 VHA specialist, BL) through identification of key concepts. Synonyms for these concepts were identified by reviewing highly cited literature, examining search terms from prior relevant reviews, considering database suggestions, and leveraging the authors’ and colleagues’ field expertise [27,33,34]. The complete search strategy is available in the Multimedia Appendix 1. The generic search query was as follows: (“virtual health assistant*” or Clinical Decision Support System or embodied conversational agent*)—AND (cancer* or neoplasm* or early detection or education or awareness or information) AND (prevention or screening) (The * or asterisks symbol is a commonly used wildcard symbol that broadens a search by finding words that start with the same letters preceding the symbol). Medical Subject Heading (MeSH) terms for the MEDLINE search on PubMed were acquired through rigorous MeSH library searches for cancer prevention and VHA. Similarly, Emtree terms on Embase were acquired using the built-in Emtree library.

Search results from each of the 4 databases were imported to the systematic review management software Covidence (Veritas Health Innovation Ltd [35], where duplicates were automatically removed by the software. A review team was set up so that each study was reviewed by 2 researchers, and any conflict over the decision was resolved by a third researcher.

The review process was conducted in two stages: (1) title and abstract screening and (2) full-text screening. The first author (AR) trained the coscreeners (XC, EGB, and LT) with the definitions of each term and the inclusion and exclusion criteria. The coscreeners reviewed the definitions and the inclusion and exclusion criteria and provided feedback on confusing terms and processes. Based on their feedback, we revised the verbiage of definitions and inclusion and exclusion criteria on Covidence. As a group, we then practice-screened 5 titles, resolved conflicts through discussion, and further revised the definitions and inclusion and exclusion criteria. Notably, 3 screeners (2 screeners and 1 tiebreaker) completed the abstract and title screening within 3 months.

For the second review step, full-text screening, AR created a hierarchy-based list of exclusion criteria on Covidence [36]. As in the first step, 2 researchers (AR and XC) screened each full-text study to make inclusion or exclusion decisions, and a third researcher (EB) acted as a tiebreaker if there were any conflicts.

For data collection, we developed a preliminary data extraction template in Covidence, aligned with our research question and standard guidelines [37]. To refine this template, the first 2 authors (AR and XC) piloted it on 3 studies[38,39,40] representing different methodologies (quantitative, qualitative, and mixed methods). After comparing their extractions, they revised the template to better focus on the research question. Both authors then independently extracted data from the remaining studies, reconciling findings through weekly discussions. Each of us then independently extracted each study and consolidated the data through weekly discussions. We collected bibliographic details (eg, author, publication date, and journal), sample characteristics, VHA features, needs identified (from “Introduction,” “Result,” or “Discussion”), and needs addressed (from “Results” or “Discussion”). When information wasn’t explicitly stated, we conducted thorough reviews before coding items as “unavailable.” The complete data dictionary appears in the Multimedia Appendix 1.

For quality assessment, we used the mixed methods appraisal tool [41], as it allows an appraisal of qualitative, quantitative, and mixed methods studies. Two researchers independently appraised each study, and consensus was reached upon discussion. Each study was assessed using 5 criteria (Table 1) drawn from the method used in Hong et al’s [41] study. These evaluation criteria were used to score each study, rating a study with 5 stars when 100% of the quality criteria are met, 4 stars when 80% of the quality criteria are met, 3 stars when 60% of the quality criteria are met, 2 stars when 40% of the quality criteria are met, and 1 star when 20% quality criteria are met (see Multimedia Appendix 1 for details).

Using our standardized template, we exported all extracted data to Excel (Microsoft Corp) for analysis. First, we synthesized VHA intervention characteristics across studies (Table 2) and participant demographics (Table 3), comparing them with each intervention’s intended audience. For thematic synthesis [30], we examined the “Introduction,” “Methods,” “Results,” and “Discussion” sections to identify audience-centered strategies (Table 4). The first author conducted initial coding of “needs identified” and “needs addressed” text segments, creating descriptive labels for audience-centered approaches. These codes were then collaboratively reviewed by the first 2 authors and grouped into thematic strategies through an iterative process, with final themes corresponding to those presented in Table 4

A search of the 4 databases yielded 1055 results, 27 of which were removed as duplicates. During the first step of screening, title and abstract screening, 893 studies were excluded as irrelevant. In the second step, full-text screening, 118 more studies were excluded for not meeting the inclusion criteria (Figure 1). The majority of the studies during full-text screening (n=94) were excluded because they were not VHA-based interventions as defined in the inclusion criteria. For the detailed screening process, please see Figure 1.

A total of 17 studies [38-40,42-55]. These studies were published between 2008 and 2022, with 6 [39,51,52,53,54,55] of them published since 2021 alone. Of the 17 studies [38-40,42-55] included, 9 [38,42,43,44,45,46,47,49,54] were quantitative, 3 mixed methods [40,51,55], 2 qualitative [39,52], and 3 multimethod qualitative studies [48,50,53]. The majority of the studies (n=11 [38,39,40,43,46,48,50,51,52,53,54] focused on colorectal cancer while others addressed prostate cancer (n=2 [42,47]), lung cancer (n=1 [49]), cervical cancer (n=1 [44]), combined prostate, colorectal cancer, and breast cancers (n=1 [45]) or general cancer (n=1 [55]; Table 1). Notably, 16 of the studies were conducted in the United States; 1 study [55] was conducted in Uganda. Two interventions were developed in languages other than English: one in Spanish [44] and the other in Luganda [55]. The majority of the studies (n=9 [39,40,43,45,48,49,51,54,55]) included both men and women in the sample, 3 studies [42,47,52] included men only and 5 studies [38,44,46,50,53] included women only. Around 6 studies [42,43,47,50,52,53] focused exclusively on African Americans, 1 study [44] focused on Hispanic women, and 1 study [55] focused on a Ugandan population. Two studies [39,40] had multiple races with a Black majority, 1 study [48] did not mention sample race. The remaining studies (n=6 [38,45,46,49,51,54]) had participants from multiple races and ethnicities in their samples, but predominantly European Americans (Table 3).

Altogether, 8 [38,40,42,43,44,46,54,55] out of the 17 studies [38-40,42-55] satisfied 3 of the 5 quality assessment criteria while another 8 studies [39,45,47,48,50,51,52,53] satisfied all 5. Notably, one study [49] satisfied 2 quality assessment criteria (Table 1). All 17 studies [38-40,42-55] satisfied the 2 screening questions pertaining to having clear research questions and a data collection method appropriate for the research questions. Hence, despite the varying levels of quality, none of the studies were excluded from the synthesis (see Multimedia Appendix 1).

Studies used inconsistent definitions of Virtual Health Assistants (VHAs), and some provided no definition at all. There was no consensus on the definition of virtual assistants, or even any standard terminology, between studies conducted by different research groups. The terms used for virtual assistants differed and included embodied conversational agent, VHA, decision aid, and other generic names such as interactive web interventions or tailored web-based interventions. The modality of the interventions evolved over time, starting with text-based web interventions and expanding by adding audio or voice and visuals for participants to interact with. Three of the interventions [38,46,55] used tailored automated voice calls as part of the intervention, and all other studies were web-based (Table 2).

Common characteristics of the VHAs in these studies included being web-based, providing tailored information based on user responses to certain questions or cues, and providing some interactivity in terms of the user being able to select responses. None of the VHAs used natural language processing. All the VHAs were preprogrammed with specific options, from which the users could select the most suitable options for themselves.

The VHAs that used animated characters for communication applied appearance-based concordance with the intended audience, with either race or gender or both (Tables 2 and 3). Notably, 8 [43,44,47,50,51,52,53,54] of the studies used animated characters to convey cancer-related information; 1 study [42] did not explicitly mention whether the characters were animated or were recordings of real humans. Specifically, 6 studies [39,50,51,52,53,54] that were all about the same intervention used an anthropomorphic name for the intervention: ALEX (Agent Leveraging Empathy for eXams) (Table 2).

Thematic synthesis of the studies resulted in 6 major strategies used to make the interventions audience-centered. These strategies can be categorized broadly into two types: (1) how the intervention was developed to be audience-centered and (2) what was customized in the interventions to make them audience-centered. The interventions were developed to be audience-centered by using 4 strategies, such as gender and race concordance, user feedback, preintervention needs exploration, and theoretical framework. Two aspects of the interventions were customized for audience-centeredness, namely information and VHA features.

The current systematic review identified existing practices in the development of VHA-based interventions for cancer prevention and screening. There are 4 key takeaways from this review. First, we found that there is an intentional effort to make VHAs audience-centered, as opposed to developing generic interventions without any audience segmentation. The studies included in this review focused on providing customized information regarding cancer and screening to the intended audience. The information was customized based on the specific cancer(s), individual risk of the cancer, guideline for screening, and screening methods. Some studies also customized different features of the intervention with surface-level cultural attributes, such as using race- and gender-matching pictures or animation and delivering information in the native language of the intended audience.

Second, the current review identified an emphasis on developing audience-centered eHealth interventions to address cancer screening barriers faced by racially minoritized groups in the United States. Around 6 [42,43,47,50,52,53] of the 17 studies [38-40,42-55] focused exclusively on developing or testing or both of VHA-based interventions to improve cancer screening among African Americans, and 1 study [44] focused exclusively on cancer screening among Hispanic American women. This is an important step toward reducing health disparities, since racially minoritized individuals often face discrimination in health care systems due to unconscious biases held by clinicians [56]. VHAs developed with the needs of racially minoritized individuals can address and mitigate such biases in preventative cancer care.

Third, of the 17 included studies [38-40,42-55], 12 [39,40,43-45,47,48,50-53,55] mentioned the importance of cultural sensitivity of eHealth interventions, and 3 [39,43,48] explicitly explained how culture was incorporated in the intervention. Rawl et al [43] evaluated the cultural appropriateness and relevance of the intervention through a community advisory board. Griffin et al [48] mentioned using cultural probing as a method for audience need assessment and the effectiveness of iterative testing of interventions among intended audiences to improve the cultural competency of the messages. Vilaro et al [39] included subjective culture and critical race theory to delve deeper into the structures of culture by exploring the individual and racial identities of the intended audience to explore the relationship between identity and perceptions of the VHA.

Studies that did not detail the incorporation of culture in the interventions or did not mention culture did not necessarily fail to incorporate cultural sensitivity. As mentioned, all studies had their interventions tested by members of the intended audience and incorporated this feedback into the intervention design. This process alone provides ample opportunity to incorporate participants’ cultural needs, even when culture is not explicitly mentioned [11]. For example, Kabukye et al [55], who conducted the study in Uganda to improve cancer awareness among the general Ugandan population, developed their intervention using interview and focus group data from Ugandan health care workers, cancer specialists, cancer survivors, and lay audience members without cancer to incorporate a multitude of perspectives in the message design.

Finally, this review identified certain areas of improvement for greater audience-centeredness of cancer prevention and screening interventions in the future. First, future research needs to identify the minoritized status of a group or individual multidimensionally (eg, including economic background, location, and education) rather than unidimensionally (eg, race only). Most of the studies in this review did not address or explore the multidimensional and intersectional aspects of being minoritized, such as income level, education level, and population location, which also affect adherence to cancer screening recommendations and screening rates [57].

eHealth literacy and education level influence individuals’ self-efficacy to use interventions, such as VHAs [58]. However, the included studies rarely recruited individuals from educationally disadvantaged backgrounds. Most of the study participants had at least a high school education. Moreover, while the majority of the studies mentioned accessibility being a barrier to cancer screening, the study participants’ location or lack of health care services in their location was not mentioned, and how eHealth technology could address such barriers was not clarified. In fact, one of the most frequently cited justifications for using VHA-based interventions is geographic accessibility. However, only 4 [39,44,53,54] of the studies explicitly mentioned recruiting rural participants. Four studies [40,42,43,55] recruited urban participants only, and the rest did not mention the rural or urban location of the participants.

In addition, studies using exclusively internet-based survey platforms or electronic health records yielded a lack of diversity in their samples [38,45,46,49,54]. Studies that specifically recruited Latina farmworkers or Black men and women used community-based recruitment or a combination of internet-based and community-based recruitment [42-44,53]. One study [39] that recruited both Black and White participants and had 55% (85/154) Black participants in the sample used community-based recruitment strategies instead of internet-based recruitment [39]. Another study recruited participants in 2 phases [51]. In the first phase, it recruited participants from a senior recreation center and recruited 24/73 (33%) Black participants and 47/73 (67%) White participants. However, for the second phase of the study, they recruited through Qualtrics (Qualtrics International Inc), with 295/1400 (21%) Black participants and 1105/1400 (79%) White participants. This finding points to a need for using multipronged strategies, or community-based recruitment, when using a single strategy to ensure equitable participation from the different racial and ethnic groups.

The second recommendation to improve audience-centered VHA development is to diversify intervention languages and implementation locations. In the current review, only 1 study [55] included a population in an African, Black-majority country, and only 2 [44,55] of the 17 studies discussed interventions available in languages other than English. The final recommendation for future researchers is to measure the social presence, that is, whether users perceived the interaction with the VHAs as an interaction with another social actor (eg, health care provider) [59] of VHAs. Though VHAs have the potential to increase social presence and thus improve patient engagement in virtual settings [25], it is rarely measured or mentioned in the review studies. Hence, in the studies included, while the communication was 2-way and immediate, whether there was a perception of social presence could not be identified. Future studies can explore how much interactivity and anthropomorphous communication VHAs can provide with intended audiences, compared to traditional or noninteractive eHealth interventions (eg, pamphlets, web pages, and emails).

The current review adds to the existing literature on digital interventions for cancer prevention and screening. Recently published systematic reviews of eHealth interventions for cancer prevention and screening have focused on specific cancers (eg, cervical cancer [60,61], colorectal cancer [62], and breast cancer [63]), specific population groups (eg, Hispanic Americans [64]), or the effectiveness of the eHealth interventions (eg, increasing screening compared to usual care [65]. The current systematic review, on the other hand, has focused on eHealth interventions’ development process and identified specific strategies used for audience-centered development. In addition, previous reviews on VHA-like interactive interventions have either focused on a specific modality of the VHA (eg, animated [34], web-based interactive tools [66], chatbots [67]). The current review included a broader description of interactive digital interventions and included VHAs of any modalities available.

Although the synthesis of the included studies in the current review was inductive, the findings can be compared to existing frameworks for developing audience-centered interventions. For example, recruiting and testing the intervention on participants who are members of the intended audience is in accordance with the principles of cultural grounding, a framework for culturally sensitive intervention development [11]. The interventions in the included studies were also customized informationally, racially, and gender-wise. Customizing has been empirically proven to be an effective way to improve health outcomes, especially among minoritized groups [9]. Finally, one of the cornerstones of audience-centered intervention development is audience segmentation, that is, developing the intervention for a specific demographic rather than a generic demographic [8]. The current review identified that the studies used some type of audience segmentation by gender, race, or age.

The systematic review provides a framework for audience-centered eHealth intervention development by identifying the strategies. In identifying these strategies, we extended our synthesis to “Introduction” or “Literature Review” sections of the studies, going beyond the “Methods” and “Findings” of the included studies, a standard practice in systematic reviews, but expanded the synthesis. This helped us identify rationales used to justify the usage of VHAs. Whether a new innovation addresses an existing problem needs to be evaluated to avoid techno-solutionism —the ideology that any societal issue, such as health inequities, is a discrete one and can be addressed by technology, without considering the multidimensional context (eg, sociopolitical, cultural, financial, and geographical) in which the issue is embedded [68]. The strategies identified in this review can be used as a framework for future intervention developers.

While the study systematically identified the major strategies used to develop VHAs for cancer prevention, we need to acknowledge its limitations. First, since there was no consensus in the literature about the definition of a VHA, the search terminologies included “decision aids” to ensure a more comprehensive result. However, based on the definition of VHA used in the study (2-way interaction and immediacy of communication), many of the decision aid–based studies were eliminated. This was because the description of the decision aid did not clarify whether the intervention had interactivity and immediacy of communication. Finally, the study included studies published only in the English language and in English-language journals. Hence, when making inferences or conclusions from this study, we must proceed with caution and acknowledge that this review is limited by language.

This systematic review establishes an audience-centered framework for developing effective VHA interventions in cancer prevention. The identified strategies, from demographic concordance to iterative user testing, provide a blueprint for creating equitable digital health tools. However, the field now faces 3 key challenges: moving beyond Western-centric models, addressing intersectional health disparities, and developing standardized measures of VHA effectiveness. Future work must prioritize interventions that account for the complex interplay of cultural, linguistic, and structural barriers to care. By adopting this comprehensive approach, researchers can ensure VHAs fulfill their potential as transformative tools in the global effort to reduce cancer disparities through technology.