Defining and Classifying CAs
All participants defined CAs as computer systems that use natural language to interact with users. In general, this definition was broad and encompassing and included voice assistants such as Siri or Alexa, which usually engage briefly with users in a “transactional” (P003) manner. At the same time, CAs often “have some kind of coherent discourse” (P003). Although some participants considered terms such as chatbot and conversational agent synonyms and used them interchangeably, most participants distinguished between them. In general, CA was regarded as a “bit broader” (P002) term that encapsulated different types of agents, including chatbots. Chatbots were seen as referring specifically to “text-based and rule-based conversational agents” (P001). However, 1 participant distinguished CAs from chatbots and other types of dialogue systems, as CAs could engage in an empathic, personalized conversation with the user and “develop relationships” (P009) that emphasized the “social and emotional aspects of the interaction in addition to the health care tasks” (P003).
Experts classified CAs according to 13 different categories. These included commonly used ones such as CA input modalities (text based or speech based), purpose (domain specific or general purpose), or response generation method (rule based, artificial intelligence [AI] based, or hybrid). They also proposed other categories; for example, CAs were classified according to the development modality as “bespoke or off-the-shelf conversational agents” (P001) or according to the deployment platform as app based, often as a stand-alone CA or integrated into a website or another platform such as Facebook Messenger, Telegram, or Slack; according to the “type of communication, is it supportive, or is it just information providing” (P006). In addition, participants also categorized CAs according to their appearance (disembodied, embodied, or social robots), length of interaction (short term, medium term, and long term), personality (coach, peer, or expert), type of CA-user interaction (transactional or relational), the inclusion of human support, “where in the patient journey it’s used” (triage, appointment management, medication adherence, others; P007), or the domain (health care, commerce, business, or others).
CAs in health care settings were seen as “a scalable way of delivering...personalized health services to people” (P010). Participants noted the advantages and disadvantages of using both rule-based and AI-based systems. There was widespread consensus among experts that, at present, rule-based CAs should be the norm in health care. Rule-based CAs require developers to predefine all dialogue turns, which allows greater control of the conversation flow, constituting their main advantage in health care settings. However, there are several disadvantages associated with rule-based CAs. For example, rule-based systems lack flexibility and may not address all user concerns. Moreover, CA interventions that require ongoing interactions with a user may lead to disengagement and boredom, as responses will be similar each time. Alternatively, AI-based CAs were seen as offering attractive, flexible, and innovative systems that may increase long-term user engagement. However, experts considered that current AI technology does not ensure safe and reliable conversations, particularly in health care settings. Finally, experts acknowledged that CA development is a complex, labor-intensive task, which applies to rule-based CAs and AI CAs.
A Conceptual Framework to Design, Develop, and Evaluate and Implement CAs
The Role of a Conceptual Framework for the Development of CAs in Health Care
Participants felt a conceptual framework to guide the development of rule-based CAs was useful. It provides researchers and developers with a guideline for CA development and “a benchmark for analysing it” (P007). Notably, a conceptual framework is helpful for “writing research proposals” (P003) or to “consult it to make sure I wasn’t missing something” (P010). Furthermore, the framework was helpful for “developers or for companies” as they may “get caught up in the business model” and “[may not] draw from conceptual or theoretical frameworks as much as they should” (P005).
All participants agreed that the framework assessed in this study could be adapted for developing AI-based CAs, as many considerations also apply to AI-based CAs. An exception was the development of the conversation flow because AI-based CAs require the use of training data sets rather than decision trees. In addition, experts pointed out that AI-based CAs’ training data should be representative of the target population, that researchers should be able to explain the algorithm’s evolution, and that the CA responses should be accurate and not compromise users’ safety.
Experts valued the content and layout of the existing conceptual framework. The visual presentation was considered “simple to understand and not overly complicated” (P008). Participants agreed to use a circular diagram as it represented the iterative nature of the CA development cycle. They also suggested the use of “non-standard fonts” (P008), “some more colors” (P006) to differentiate the stages, “emphasizing the arrows” (P008) to highlight the framework’s iterative nature, and modifying the placement of the cross-cutting themes to avoid linking them to the specific stages of the framework. However, they suggested adding more details to each stage to make the diagram more self-explanatory. Participants had varied views about the framework name, appreciating that it was simple and easy to remember, although they might “not directly connect it with the framework for chatbot development” (P002).
Design of CAs
Participants agreed with the following existing concepts: defining the CA goal, determining the CA identity, identifying target users, selecting the delivery interface, and assembling a multidisciplinary team. They also added a new concept: specifying the evaluation outcomes (Multimedia Appendix 2). Defining the CA goal was highly relevant. Participants expressed that a clear and in-depth understanding of the health care problem is essential to “define the goal based on the problem” (P001). Still, they highlighted the relevance of defining “the central aim, why are we doing this?” (P005). Research methods used for needs assessment, such as reviews, focus group discussions, interviews, and surveys, play an important role not only in defining the goal but “for all of the content to make sure it’s understandable and acceptable by [the] end audience” (P003).
The identity of the CA is important “in terms of who people trust and don’t trust and the degree of which they think the information they’re getting is trustworthy” (P009). One fundamental aspect of defining the CA identity design is to adapt the content to the local culture, including the local nuances that may make the agent more relatable to the target audience.
The target users refer not only to the end user of the intervention but also to the broader social circle formed by family, other members of the user’s social network, and health care providers. This is particularly important if the CA intervention will be used in the health care system or if the target population includes older adults or individuals with intensive care needs. Developers may also consider the setting where the intervention would occur.
Selecting the delivery interface is “a key choice” (P012). Researchers “need to realize that there are differences with the platform” (P010), in terms of users’ acceptance of the different tools, and regarding the technical “capabilities of the delivery interface” (P008). The choice of platform may even determine the type of outcome measures to be collected during the CA evaluation.
Assembling a multidisciplinary team was considered essential when designing a CA-based intervention. The team may include a “linguistic professional” (P012) to ensure that the text will be adequate for the end user, content participants familiar with traditional face-to-face interventions, and digital health intervention specialists. The team composition may also depend on the nature of the research collaboration, as 1 expert pointed out, “if I'm working with [an industry stakeholder], they tell me the team” (P008). Finally, participants suggested that it is also important “to define the roles of everyone within the team because sometimes that’s a bit unclear” (P007).
Experts noted that the type of outcomes to be measured and the time points at which these measurements may occur during the evaluation stage should be defined during the design stage. It is crucial to consider the sources of “the data that you’re going to use...it’s participant reported and sensor data, but also electronic medical record data, health systems data, staff data, leadership data, focus group data. So, it’s broader than just participant data” (P009). Planning for sensor data collection is essential as it must be coded while developing the CA. Special care should be taken to ensure the correct spelling of data variables, as minor orthographic mistakes may render the data unavailable.
Development of CAs
Although participants agreed on the concepts included in this stage, they suggested adding additional details to the diagram to make the content more self-explanatory, particularly for inexperienced developers.
The length of the dialogues was repeatedly mentioned as a critical consideration to be discussed early in the development process. It may take different interpretations, referring to the length of each CA turn (how many lines of speech), the length of the dialogue (number of times each speaker has a turn), and the total number of sessions required to deliver the intervention. The dialogues may convey an empathic, nonjudgmental tone, as this may influence user engagement and bonding with the CA. Moreover, it is essential to consider the language that best translates the content of face-to-face interventions into the constrained length of CA dialogues.
Participants agreed that all content in the conversations should be evidence based, referring to the adherence to current best practices as stated in the scientific literature and clinical guidelines and the dialogue structure that should emulate “seasoned clinicians” (P009). Health care CAs must also have adequate systems to respond to possible medical or psychological emergencies. For example, they may provide crisis helpline phone numbers, include advice for the initial management of common symptoms, or hand over emergency management to a health care provider.
Error management includes inadvertent adverse events and mistakes (either from the CA or a person who enters the wrong data). However, adverse events are rarely reported in CA studies, which increases the importance of collecting “information if a person deteriorates in an outcome” (P006). Moreover, 1 expert highlighted that adverse event management might be challenging if users associate the CA with their health care provider and assume that the agent would identify adverse events when they occur. Therefore, developers may include safeguards to assist users adequately if an adverse event occurs.
Personalization implies that the “chatbot can maybe store the user preferences during the conversation” (P011). Personalization is essential, and it is considered a hallmark of CAs, which may draw information from several sources to “tailor the dialogue specifically to [the user]” (P009).
Finally, when defining the platform, researchers should consider the size and expertise of the multidisciplinary team. Although large teams may develop the CA “from scratch” (P001), smaller teams with limited coding expertise may “use some platforms, [such as] the Alexa skills or Juji” (P001).
Evaluation and Implementation of CAs
Participants agreed with existing evaluation and implementation topics, such as usability assessment, user engagement, intervention efficacy, and effectiveness, and suggested new topics, such as technical evaluation. They also pointed out that CA evaluation is an iterative process that often starts during content development to ensure that conversational turns occur as planned and are adequate to fulfill the CA aim. In addition, during the evaluation process, researchers should consider how the test results can be used to improve the CA further.
Participants offered varying definitions of what usability entails. For example, usability was defined as “not just A/B testing, but qualitative studies of what users think about the agent” (P003) and referred to user experience as a dimension of usability. Concurrently, it is important to use standardized tools to measure usability, such as the System Usability Scale [41] or the Chatbot Usability Scale [42], as “researchers tend to develop their own usability test” (P002), and this may limit comparability. User engagement is another important consideration, particularly in longitudinal interventions.
The technical evaluation of CAs has not been discussed in the original framework. However, such evaluation is important to ensure that the CA is ready and functional before evaluating it in clinical trials, probably during development. It includes not only crashes and bugs but also the evaluation of the CA content credibility and data privacy and security safeguards, among others.
In general, participants agreed that, although infrequently mentioned, economic evaluation of the end product is essential to ensure the viability of the CA beyond clinical trials and consider the cost-effectiveness of the intervention. Moreover, expert P002 suggested that economic evaluations could be considered in all 3 stages of the CA development process, whereas expert P003 argued that “economic evaluation is rarely done” in research settings.
Finally, 1 expert suggested modifying the framework to “having four boxes and four quadrants” as implementation is different from evaluation and requires “different skill sets altogether” (P009).
Cross-Cutting Themes
Participants agreed with the existing cross-cutting themes of user-centered design and data privacy and security and suggested 2 additional ones: ethics and long-term sustainability.
Including user-centered design was an important aspect of CA design and evaluation. However, participants clarified that user-centered design is the name of a specific design process characterized by the user’s influence in the design and should not be used to conceptualize the importance of user involvement in the design process. Participants distinguished between being user centric; seeking end users’ opinions through surveys, interviews, or other qualitative methods during the CA development; and directly involving users in the design and development of the CA, as seen in participatory design frameworks. However, users may express divergent views, which may be difficult to reconcile when deciding the features to be included in the intervention.
The inclusion of data privacy and security was welcomed by all participants, who agreed that this topic is critical and sometimes not adequately discussed in the scientific literature. Data privacy and security were perceived as complex issues with multiple connotations, including users’ behavior, which could be “very paradoxical” (P008), meaning that users may assert themselves as genuinely concerned about privacy but act as if it is not important at all. Users’ concerns about data management may correlate with the extent of personal data collection. They may require more detailed explanations of how researchers will use their data if the app collects large amounts of personal data. In addition, researchers need to find a “balance between getting as much data as possible... [to] make the intervention more personalized, but also [not] to cross the boundaries of what people feel comfortable with sharing” (P007). Participants also emphasized the importance of developing adequate data management plans that align with their countries’ current data protection laws during the design stage. Finally, future handling of research data may be challenging, as technology development may facilitate the reidentification of initially anonymous data.
Ethical considerations were viewed as an essential cross-cutting theme from 2 different perspectives. One expert referred to the overarching ethical principles that should guide all health care interventions, “principles like nonmaleficence, beneficence, autonomy, fidelity” (P005). In contrast, another expert mentioned that when applying for ethics approval, researchers may need to modify the project execution to suit the current ethical best practices.
Finally, expert P010 suggested assessing early in the development (and as an ongoing theme) “where [the intervention] fits in the bigger health system” to ensure that researchers develop a sustainable, cost-effective system that addresses a real and required health care need.