Conversational agents (CAs) are intelligent computer programs empowered with natural language processing techniques that engage users in human-like conversations to provide an effective and a smart communication platform in a simulated environment, including text-based chatbots, voice-activated assistants, and embodied CAs [1,2]. They are designed to obtain specific information from users that is necessary to perform particular tasks and respond in a manner that is optimal to achieve these goals. Due to their ability to transform the health care system and enable individuals to comanage their health care effectively, CAs are increasingly used to deliver health care services [3]. The most popular health CAs include ELIZA [4], Casper [5], MedChat [5], PARRY [6], Watson Health [7], Endurance [7], OneRemission [8], Youper [9], Florence [10], Your.Md [11], AdaHealth [12], Sensely [13], and Buoy Health [14], among many others. CAs are being tested and adopted to provide and collect health-related information and provide treatment and counseling services [15]. In some cases, they are used to enhance the accessibility, efficiency, and personalization of service delivery and ensure relatively equal delivery of health care services worldwide through bridging the gaps between developing and developed countries [15,16].

Given the growing significance of CAs, researchers have conducted a plethora of relevant studies, varying from their suitability as health care partners to their designs including physical appearance, gender, and speech. [17-20]. These studies aimed to improve “humanness heuristics,” affective states in users, and user perceptions of the CA personalities by tailoring CAs to the cultures and demographics of the users to continuously promote user engagement, adherence, and adoption [21-24].

Language plays a crucial role in improving user engagement because perceived impersonal closeness, intention to use, user satisfaction, establishment of trust, and user self-disclosure or self-concealment are closely associated with the task- and social-based interactivity, interaction, politeness, and information quality provided by CAs [2,19,21,25-32]. However, few studies focused on language use in CA-based health communication to examine its influence on the perceived usability of CAs and the perceived roles of CAs in delivering health care services [16]. Language considerably influences the joint construction of meaning between interlocutors and rapport establishment [23,33,34]. This is particularly true for human-machine communication. For example, when addressing users by their first names, CAs are perceived to display varying degrees of politeness and thoughtfulness determined by cultural limits and preferences [24]. It follows that intensive and extensive investigations into language use by CAs in different linguistic settings are crucial to scale up health care interventions delivered by CAs worldwide [16,35].

The language use of an information source is likely to be crucial among various factors affecting the information seekers’ judgments on the credibility and trustworthiness of the information providers [36-39]. In this review, language use, characterized by various linguistic aspects, is defined as varied verbal strategies and compliance-gaining techniques [40] that the CAs under scrutiny adopted to deliver health interventions. These strategies and techniques may involve various ways of wording, including an everyday style (eg, “heart attack”) versus a technical style (eg, “myocardial infarction”), a tentative style (eg, “presumably similar”) versus a nontentative style (eg, “similar”), a neutral style (eg, “methodological mistakes”) versus an aggressive style (eg, “really dumb methodological mistakes”), an emotional style versus a nonemotional style, and an enthusiastic style versus a nonenthusiastic style. [36,41-44]. They may also include the use of personal references (eg, first-person and second-person pronouns), personal testimonials, specific conversational frameworks or prompts, and other verbal means of communication [45-47]. In short, the language use of the CAs under discussion in this review refers to their characteristic linguistic performances in health communication.

Language Expectancy Theory [48] and Communication Accommodation Theory [49] assert that acquiring knowledge when seeking web-based health information is determined not only by the information content but also by who is communicating the information and the manner and context of communication. Information seekers evaluate information providers positively if the latter’s language use is in tune with their cultural values and situational norms and if they use language more favorably than expected in a situation [50]. The language use of information providers is regarded as a prominent clue to evaluate the characteristics of the providers, especially in web-based communication [51,52]. The information provider’s language use is a cue for determining whether people perceive the information to be credible and whether the information provider is trustworthy [37,46].

The current review aimed to summarize the language use of CAs in health care to identify the achievements made and the breakthroughs to be made to inform researchers and more particularly CA designers and developers. This can help realize the high potential of CAs for improving individual well-being.

The primary objective of the current review was to identify the language use of CAs in health care. This review was performed by following the protocols of the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement [53]. We first designed a search strategy according to the research aim and then performed keyword searches in PubMed and ProQuest databases for retrieving relevant publications. Then, 3 researchers screened and reviewed the publications independently to select studies meeting the predefined selection criteria. Finally, we synthesized and analyzed the eligible articles.

This review focused on two aspects of the previous studies: CA applications in health care and language use. To retrieve a high number of relevant studies, we decided on using the keywords relating to language use for literature search, including “expression,” “language,” “language style,” “language feature,” “language characteristic,” “language pattern,” “linguistic style,” “linguistic feature,” and “linguistic characteristic.” Based on these keywords and those concerned with CA applications in health care, we developed the following search strategy to identify studies wholly or partly investigating the language use of CAs: ((expression [Title/Abstract]) OR (language [Title/Abstract]) OR (language style [Title/Abstract]) OR (language feature [Title/Abstract]) OR (language characteristic [Title/Abstract]) OR (language pattern [Title/Abstract]) OR (linguistic style[Title/Abstract]) OR (linguistic feature [Title/Abstract]) OR (linguistic characteristic [Title/Abstract])) AND ((health* chatbot [Title/Abstract]) OR (health* conversational agent [Title/Abstract])). Drawing on this search strategy, we conducted keyword searches in 2 databases (PubMed and ProQuest) to retrieve published papers without restrictions regarding the year of publication on February 11, 2022.

We included both peer-reviewed and non–peer-reviewed journal publications because the aim of this review was to provide a comprehensive overview of the language use of CAs in health care and its corresponding implications for improvement in language use in CA communication to inform future research and CA designers. Textbox 1 shows the inclusion and exclusion criteria.

We used Microsoft Excel (Microsoft Corporation) to manage the collected articles by listing the titles, abstracts, and article types for screening. First, 2 researchers (MJ and YS) screened the titles and abstracts of the candidate articles independently, filtering those articles that did not conform to the selection criteria. If the eligibility of some studies was unclear, we included them for further full-text review. Then, 2 researchers (YS and WX) reviewed the full texts of the remaining articles independently. Any disagreements were resolved through discussion and consultation with the third researcher (MJ).

To analyze and synthesize the language use of the health care CAs, the following information was extracted from eligible studies by YS: first author, year of publication, health care application, target population, study design, major findings, and limitations. Then, MJ reviewed and cross-checked the extracted data. Any discrepancies were resolved through a discussion with the entire research team.

A meta-analysis was not feasible due to the expected variety of health care applications, target populations, study designs, results, and limitations. Therefore, we conducted thematic synthesis to summarize the data extracted from the included articles following 3 steps, namely “line-by-line” coding of the text, development of “descriptive themes,” and generation of “analytical themes” [54]. YS first coded each line of the extracted text according to its meaning, then developed descriptive themes, and finally generated analytical themes using the derived descriptive themes [55]. MJ validated each assigned code, each derived descriptive theme, and each developed analytical theme independently. All the authors discussed and finalized the results of the thematic synthesis.

Using the search strategy, we identified 179 publications in the PubMed and ProQuest databases. From these retrieved publications, 40 were eliminated because they were not journal articles but were other types of publications (eg, commentaries, letters, news, and editorials); 51 were eliminated for being duplicates, and 72 for not meeting the selection criteria. After the full-text review, another 5 studies were excluded; 3 were not related to language communication, 1 was not about health care, and 1 was an editorial. As a result, 11 studies met the inclusion criteria and were eligible to be considered in this systematic review. Figure 1 shows the screening and selection process.

Table 1 summarizes the information extracted from the 11 papers selected for synthesis and analysis. The major findings reported in the original studies that are directly related to the aim of the current review are included. We have included the limitations reported in the original studies and those based on our perspectives, if any. Based on this table, we present the qualitative synthesis and analysis in the Discussion section.

In human-CA linguistic communication, language cues are particularly important because they perform a crucial function in promoting user engagement [2], but few studies examine significant sociolinguistic dimensions in CA design across different languages and cultures, and the impact of these dimensions on user perceptions of CAs and their effectiveness in delivering health care services [16]. In this review, 6 of the 11 included publications deal exclusively with the language use of the CAs studied, and the remaining 5 are partly related to this topic. We derived the following themes from the language use in the 11 included studies through thematic synthesis.

Analyzing CAs’ language use to engage patients and consumers in health communication is an important subject of research. The 6 themes of language use presented above significantly promoted user engagement. Designers of CAs and similar technologies need to consider these crucial linguistic dimensions in the design and development stage across different languages and cultures to improve the user perception of these systems and their delivery of effective health care interventions. Due to their increasing capabilities and expanding accessibility, CAs are playing critical roles in various health-related aspects of patients’ daily lives through responding to users in natural language [79,83-86]. Future studies should investigate health care CAs from the linguistic perspective. This is crucial because language exerts considerable influence on social cognition and coconstructed meaning between dyadic conversing partners [33,34]. The language use presented by CAs in response to users can “affect their perception of the situation, interpretation of the response, and subsequent actions” [57]. Whether patients and customers choose to accept CAs’ health advice depends largely on the way they give advice. Good advice is judged by the advice content and its presentation [87]. “Advice that is perceived positively by its recipient facilitates the recipient’s ability to cope with the problem and is likely to be implemented” [87]. Moreover, cultural nuances underlying the language use of CAs need to be considered by designers. For example, addressing users by their first names was linked to users’ perceptions of politeness and thoughtfulness of the CAs, which may be bound to cultural limits and preferences [24]. Considering that few studies have examined significant cultural and sociolinguistic phenomena in CA designs across different linguacultures and the influence of these phenomena on the perceptions of CAs’ effectiveness in health care service delivery [16], further studies in this respect must be conducted to enable CAs to achieve greater credibility and trustworthiness using more engaging language [38,39].

Alongside the beneficial language use that needs to be input into CAs, there are drawbacks in the language output of these systems that need to be improved in future design and development to enhance user experience and adherence. Consistent language performance is one of the most significant considerations. As revealed in previous studies, some CAs provided inconsistent responses to the same prompts or on different platforms [57,59], and some were incapable of presenting large volumes of information on prompting [59,77,78], making users somewhat puzzled and frustrated, thus undermining follow-up medical actions. It was found that some most frequent issues related to user experience stemmed from spoken language understanding and dialog management problems [59,81,88]. Although CAs capable of using unconstrained natural language input have gained increasing popularity [89], CAs currently used in health care lag behind those adopted in other fields (eg, travel information and restaurant selection and booking), where natural language generation and dialog management techniques have advanced well beyond rule-based methods [90,91]. Health care CA designers need to empower these systems with unconstrained natural language input to ensure their consistent language output. Moreover, advances in machine learning, especially in neural networks, need to be integrated into the design of CAs to empower these systems with more complex dialog management methods and conversational flexibility [92,93].

Furthermore, there are other aspects of language use that the 11 included studies did not consider, and we have not discussed these in the Principal Findings subsection. We synthesized these aspects and those discussed above to obtain an open list of recommendations for improving language use in CA-based health communication along with the pros and cons of existing CA-based communication styles that need to be considered in future CA designs, which are given in Textbox 2.

This systematic review has some limitations. The first one was attributed to the retrieval of relevant articles. We searched PubMed and ProQuest for suitable publications. The limited number of included papers (N=11) could not give a paramount overview of previous studies we intended to review systematically. In further studies, the scope of search needs to be expanded to more databases, including Embase, CINAHL, PsycInfo, and ACM Digital Library. Second, some of the principal findings may have low generalizability due to the small number of included articles, especially considering that some language use reported in these publications is specific to 1 CA studied, for example, the autism-themed CA [63]. Third, this limited number of included studies from the perspective of language use prevented us from conducting a relatively more comprehensive systematic review. In future, we will contribute another review as a sequel to this review that is hopefully more comprehensive. Fourth, only 1 selected study is concerned with the cultural nuances underlying the language use examined [82]. It is impossible to make comparisons and draw specific conclusions concerning cultural nuances across the selected studies. This is a limitation that needs to be overcome in future research.

Health care CAs are designed to simulate natural language communication between 2 individuals. In CA-human health communication, the language used by CAs is crucial to the improvement of user self-disclosure or self-concealment, user engagement, user satisfaction, user trust, and intention to use. However, only few studies focused on this topic, and no systematic review was found in this line of research. Our review fills this gap in the literature. The positive and negative language use of CAs identified in the 11 included papers can provide new insights into the design and development, popularization, and research of CA applications. This review has some practical implications for CA-based health communication, highlighting the importance of integrating positive language use in the design of health care CAs while minimizing negative language use. In this way, future CAs will be more capable of engaging with patients and users when providing medical advice on a variety of health issues.