In 2016, noncommunicable diseases (NCDs) accounted for 40.5 million deaths worldwide, which corresponded to 71% of deaths worldwide. The top 4 NCDs are cardiovascular diseases, cancers, diabetes, and chronic lung diseases [1]. NCDs can be prevented by adopting a healthy lifestyle. For example, a European multicohort study, which was conducted from 1991 to 2006 and included 116,043 people who were free of major NCDs at baseline, suggested that various healthy lifestyle profiles yield gains in life years without major NCDs, including type 2 diabetes, coronary heart disease, stroke, cancer, asthma, and chronic obstructive pulmonary disease [2].

Changing and sustaining health behaviors, which are integral to both the prevention and self-management of NCDs, are known to be challenging and resource intensive [3]. Digitalization and automation remove time and place restrictions, thereby broadening access to lifestyle and self-management interventions in a potentially cost-effective manner. For example, full economic evaluations of interventions that use the internet, mobile devices, or computers for the prevention and control of type 2 diabetes have demonstrated high cost-effectiveness, even though they were not fully automated [4].

The European Blueprint on Digital Transformation of Health and Care for the Ageing Society reflects a common vision that key stakeholders have on the role of innovation in changing health and care provision among older populations [5]. The priority topics encompassed by this policy vision are disease prevention, personalized health and care, and digital tools for citizen empowerment and person-centered care. A total of 12 blueprint personas have been created based on the health and care needs of people across the life course, ranging from children to persons aged ≥80 years [5].

Digital technology that mimics human communication is suitable for different age populations and populations with different literacy levels and is arguably more engaging for long-term use. A scan of this landscape has revealed a considerable body of scientific literature, although no agreements have emerged on the definition of so-called conversational agents. For instance, some authors consider conversational agents to be software capable of natural language processing [6]. However, others have used broader definitions that encompass agents that use predefined text options as inputs but exclude embodied agents that use nonverbal communication [7].

For the purpose of our review, we defined conversational agents as computer programs that are designed to simulate two-way human conversation by using language (speech or text) and are potentially supplemented with nonlanguage modalities. We believe that conversational agents for health and well-being should be further characterized according to the health intervention (eg, target population, design, the entity on which the intervention is carried out, and duration), the agent (eg, embodiment, role, and delivery channel), and the conversation (eg, input and output options, dialogue engine, and sentiment detection). We derived these ideas from literature [7-11], the international classification of health interventions [12], our interdisciplinary experience [13-16], and discussions within the research team.

This study is part of the Harnessing the Power of Conversational e-Coaches for Health and Well-being Through Swiss-Portuguese Collaboration (eCCo) project [17], which encompasses an evidence map and the subsequent development of an international taxonomy on conversational agents for health and well-being via a scientific consensus method that will be informed by our literature review. In addition to this purpose, the evidence map independently serves a much-needed research endeavor—fostering collaboration in the field through an open-access resource. In their review on conversational interfaces for health, Xing et al [18] highlighted the need to improve collaboration among stakeholders in research and patent activities. To our knowledge, there is no open, searchable database on conversational agents for health—a resource that could foster collaboration by pinpointing expertise across international centers and networks. Such collaboration can help with tackling research fragmentation and duplication.

Our aim is to provide an overview of all relevant evidence on conversational agents for health and well-being across the life course. Specifically, our objectives are to identify, categorize, and synthesize primary studies and reviews on this topic by focusing on the following research questions:

The literature search in MEDLINE; CINAHL; Web of Science; Scopus; the Cochrane, ACM, IEEE, and JBI databases; and Google Scholar was conducted between November 11 and November 19, 2020. A total of 16,351 records were identified and exported to Rayyan. The removal of duplicates yielded 8022 records, which were subjected to screening (Figure 1).

As of April 2021, we produced an operational procedure to support screening (aided by Rayyan) to ensure consistency and reduce the amount of errors. We also drafted a glossary with operational definitions to support the study selection and data extraction procedures. This is regarded as a living document, which will be updated as our work progresses (Multimedia Appendix 3; the glossary is currently based on 5 publications [7,45-48]). Moreover, we concluded the first the stage of screening, as depicted in Figure 1.

Our work aims to provide an overview of all relevant evidence on conversational agents for health and well-being across the life course. The evidence and gap map is a type of literature review that is particularly suited for this aim. As White et al [50] put it, “[evidence and gap maps] usually show what evidence is there, not what the evidence says.” By following the same systematic approach as a systematic review, we aim to offer a wider picture of the research landscape by including both primary studies and meta-research. To the best of our knowledge, published reviews on conversational agents, such as those of Car et al [7] and Chattopadhyay et al [27], have only included primary studies. Therefore, we add to current knowledge by extending the scope of existing reviews. As for other evidence maps, the novelty of our evidence synthesis comes from its breadth (ie, summarizing all reviews and primary studies on the topic without collating effects or effect sizes), while the strength of the scoping reviews comes from their depth (ie, typically a narrower scope and direction of effect) [50]. Another aspect regarding the broader scope of our work is our definition of a conversational agent, which comprises not only chatbots but also agents with physical or web-based embodiments, such as robots or anthropomorphic web-based agents.

The number of reviews that were preliminarily uncovered by our review also suggests that it may appropriate to conduct an umbrella review of meta-research in the future. This will help those aspiring to conduct a review in this field to avoid wasteful redundancies. The number of reviews we uncovered raises the issue of the degree of overlap and the incremental value of these publications. Recently, Tugwell et al [51] elaborated on the replication of systematic reviews. In addition to direct replication, which involves repetition for verifying results, these authors put forward the concept of conceptual replication, in which a research question is broadened or narrowed to ascertain different intervention types, settings, outcomes, or study designs [51]. The data extraction of the included reviews in our work will clarify whether these reviews performed conceptual replication or undesirable repetition, which has been coined as research waste [51]. We envisage that the open, searchable database that will be developed in the data synthesis stage will help researchers to avoid future research waste by highlighting published reviews.

Another methodological consideration is the role of bias and its influence on the evidence map results. For instance, we addressed bias in the selection of studies through a multifold procedure [52]. First, we detailed the research questions and the inclusion and exclusion criteria to avoid inconsistent application. Second, we developed a glossary of operational definitions to reduce discrepancies in the interpretation of key terms. Third, per the review protocol, two independent reviewers will screen and extract data, and a third reviewer will be involved when discrepancies cannot be resolved by consensus. Other procedures specified in the protocol include presenting a PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram for results, pilot testing the several steps of the review method (searching, selection of studies, data extraction and critical appraisal), and having workflows subsumed in the evidence map.

The evidence and gap map relies on a framework for detailing its dimensions, which will be operationalized by using row and column headings [50]. The eCCo project will tackle this issue by pursuing an integrated approach based on the interdependency of its two core activities, as follows: (1) producing an evidence map on conversational agents in health and well-being and (2) consensualizing a taxonomy on the topic. The matrix for the data extraction (Multimedia Appendix 2) was informed by the first draft of this taxonomy, and the evidence map will allow for the identification of international experts who will be involved in the consensus conference. Further, the results of the evidence map will be used to fine-tune the taxonomy draft before the draft is subjected to the scrutiny of international experts. Our research project is in line with the work of Bittner et al [53], who used an empirical-to-conceptual approach for the development of a taxonomy for conversational agents and drew upon a literature review to identify new subsets of objects. We identified a set of taxonomies in the field, albeit none were health specific [8,9]. The foci of all these taxonomies are design options that do not detail the aspects of the health intervention. Guidelines for reporting on digital health interventions [54,55] have recommended the specification of intervention components and modes of delivery (eg, specifying who delivers the intervention, who receives the intervention, how often the intervention is delivered, the intervention duration, the format of the intervention, and the context in which the intervention is delivered). These requirements were considered when first drafting the eCCo taxonomy.

None of the above-mentioned taxonomies used a scientific consensus process to standardize terminology and categories of conversational agents. This is a limitation that we are addressing via the eCCo project.

In spite of its clear strengths, the review will not be without limitations. Integral to evidence maps is the fact that study outcomes are not extracted to ascertain effects. Moreover, the fact that we will not appraise the quality of primary studies means that we cannot pinpoint research gaps for areas with high volumes of potentially poor-quality studies. Another limitation is that research gaps do not necessarily translate to research needs; when prioritizing research needs, one should consider aspects such as relevance and potential impact. Nonetheless, synthesizing evidence on health-focused conversational agents will facilitate the prioritization of strategic research by commissioners.

In addition to fostering collaboration, we envisage that the open, searchable database will also contribute to bridging the translational gap by, for instance, identifying projects with a higher technology readiness level. Such projects can more easily reach the market via partnerships with the business sector.