This systematic review of the English-language scholarly literature followed standard guidelines for conducting and reporting systematic reviews, including Preferred Reporting Items for Systematic Reviews and Meta-analyses [41] and guidelines from the Cochrane Diagnostic Test Accuracy Working Group [42,43].

Literature searches were performed from inception to December 31, 2019 in Google Scholar and 7 online databases: MEDLINE, EMBASE, PsycINFO, CINAHL, Cochrane Central Register of Controlled Trials, PubMed, and ACM Digital Library. Search queries covered 3 domains: (1) avatars, (2) (digital) narratives, and (3) virtual assistants (for details, see Multimedia Appendix 1). Our search was limited to peer-reviewed articles published in English. Manual searches were extended to bibliographies of review articles. Figure 1 shows a summary of the literature search.

The topic of our review crosscuts two disciplines — health care and computer science — and their disciplinary priorities are divergent if not orthogonal. On the one hand, health care research prioritizes reporting and replicating empirical evidence of efficacy. On the other hand, driving innovation is a key mission of computer science research. Thus, empirical investigation of these innovative designs — particularly replication of such studies — often takes a back seat. To offer a comprehensive review of the topic at hand, we first present a qualitative, systematic review of VH in patient-facing systems. Some of those articles were then included in a quantitative meta-analysis (Figure 1).

Studies were included in the qualitative review if they met the following criterion: designed or evaluated VH for health-related outcomes in a patient-facing system. Studies were excluded if they used VH for the training and education of health care professionals or students.

Some of the articles included in the qualitative synthesis were further included in a meta-analysis if they met the following criteria: (1) compared the effectiveness of VH in a health-related outcome in a target population against a control group with no VH; (2) studied humans of any age; and (3) reported the sample size and mean and variance of the outcome measure in control and experimental groups. Studies were excluded if they did not use a comparator that was equivalent but different from VH (eg, [44]).

All records were first downloaded into an EndNote X8.2 library [45], and duplicates were removed. Titles and abstracts were then screened for inclusion and exclusion criteria. To accelerate screening, records were first collated into topics using keyword searches in the EndNote library and then reviewed for inclusion. Two investigators (the first and second authors) independently assessed the relevancy of search results and selected full text articles for further review. The second and third authors independently abstracted the key study factors into a data extraction form and then came to a consensus on which studies met the inclusion criteria; the first author made the definitive decision when discrepancies would arise.

Data from eligible articles were extracted into a spreadsheet. For the qualitative review, data included target population, design objective, type of evaluation, principal findings, and VH characteristics. For the meta-analysis, studies were summarized using the PICOTS (population, intervention, comparison group, outcomes, time frame, setting) framework [46].

Articles included in the meta-analysis were assessed for quality and risk of bias using the latest criteria from the Cochrane Consumers and Communication Review Group [47]. Studies were categorized as low risk, unclear, or high risk for each Cochrane bias domain: selection, performance, detection, attrition, reporting, and other [47]. Publication bias was assessed using Egger’s test [48] and a P-curve analysis [49] and graphically examined using funnel plots [50].

All statistical analyses were performed in R using the meta [51] and metafor packages [52,53]. Standardized mean differences (SMD) were used as the effect size to quantify the effectiveness of VH. The overall effectiveness was estimated using a random-effects model, and forest plots were used to graphically present the combined effectiveness. Statistical heterogeneity among the studies was assessed by I², which estimates the percentage of total variation across studies due to heterogeneity rather than chance alone. We considered heterogeneity statistically significant at P<.05 and used random-effects models to take into account the heterogeneity among the included studies. Subgroup analysis was conducted across studies by the type of health outcome — health and wellbeing as well as attitudes toward health and wellbeing.

A total of 16,794 search records were retrieved from the databases, and 1,985 additional records were identified from the bibliographies and Google Scholar. After removing duplicates, we screened the titles and abstracts of 8,125 articles; 380 articles involved a functioning VH system outside a game. Because the computing technology for creating VH did not exist prior to circa 2000, all studies published before 2000 were excluded. These 380 abstracts were further reviewed for their relevance to health, and 282 articles were excluded because the VH was involved in contexts like training, education, or demonstration. The remaining 98 articles underwent full-text review, and 53 articles met the inclusion and exclusion criteria for the qualitative systematic review.

A total of 30 health-related outcomes were identified in the 53 eligible articles, targeting 25 types of populations (Table 1) and 6 modalities of technology delivery apart from a personal computer, desktop, or laptop (Table 2). Unconstrained speech input was sparsely used in VH systems [54,55]. While most systems allowed constrained user input via a touchscreen or keyboard, some systems reacted to nonverbal conversational inputs, such as gaze [56] or proximity [57].

In the 53 eligible articles, 33 unique systems were identified. Of these, 9 systems were used for health assessments and the rest for health interventions (Table 3). Two broad design categories emerged — simple VH and VH with health sensors or trackers (Table 3). These additional trackers did not augment the interaction capabilities of VH, but provided additional data about users (eg, via a heart rhythm monitor, respiration sensor, or eye tracker).

A total of 26 studies (16 articles) published between 2000 and December 31, 2019 were eligible for the meta-analysis, targeting 11 types of populations and including 10 studies with healthy adults [68,69,81,83,87], 3 studies with older adults [63,67], 3 studies with women [58,88], 2 studies with individuals with schizophrenia [78,95], and 3 studies with individuals with depression [72,76].

This meta-analysis found VH interventions significantly more effective than other types of traditional interventions that did not include conversational agents (SMD .166, 95% CI=.039-.292, P=.012). This is an important finding because effective interaction modalities are needed to promote improved consumer or patient engagement in health, thereby promoting behavior change and the management of chronic conditions. The effect size may have been small because the studies involved many different population and outcome types. As more studies become available in the future, some populations and outcome types may show stronger effect sizes than others.

The efficacy of VH may also depend on the type of intervention for which they are used, such as delivering cognitive behavioral therapy vs delivering education. However, the number of current studies was insufficient to conduct such subgroup analyses. Nevertheless, the effectiveness of VH did not significantly differ between health-related outcomes and health-related attitudes.

Like prior reports [37], we found considerable heterogeneity in evaluation methods and outcome measures and a predominance of quasiexperimental study designs or usability studies over RCTs. But unlike previous reviews [36], we found most applications were fully developed, and many of them were either evaluated or soon to be evaluated in RCTs. Our findings cannot be compared directly with the previous reviews because each of the reviews used a different definition of conversational agents or avatars. Typically, earlier reviews did not discriminate between conversational agents with and without a humanlike physical appearance; therefore, they did not discriminate between whether they were able or unable to engage in nonverbal conversation. Only 5 of the 26 studies that were included in the meta-analysis compared the efficacy of VH interventions with text-based chatbots or any other type of conversational agents (Multimedia Appendix 2) [58,68,87,91,92]. Other studies compared interventions against treatment as usual, such as a therapist visit or standard educational materials. Voice-based and text-based conversational agents, without a physical humanlike appearance, were used as comparators in 2 studies [58,92] and 4 studies [68,87,91,92], respectively. Thus, it is too early to tell how much the efficacy of VH in health-related interventions can be attributed to their humanlike physical appearance.

Of course, some health applications, especially some mental health assessments, would only work with VH and could not be replaced with a voice-based or text-based conversational agent. For example, VH were used to assess the influence of unusual voices on daily activities of hallucinating patients [74] and the effect of social rejection in individuals with psychotic disorder [98]. But apart from these, when should a VH intervention be more desirable than other types of conversational agents? Engineering health-related VH interventions is far from trivial and may further cause unintended effects, such as a lack of empathy or a sense of complacency among users [23,24]. When is it worth investing in designing VH over chatbots or smart speakers? This question kept resurfacing in our analysis but remained unanswered. As more studies become available, subgroup analyses of specific interventions could answer this question.

Currently, the prevalence of VH in health applications appears to lag behind that in other areas, such as education and training [118,119] or serious games [120], which is not surprising given the concerns about patient safety [121], ethical and legal issues [122], and perception of professional counselors [123]. Technological advancements will continue to augment VH capabilities; equally important is identifying the design tradeoffs associated with those capabilities in patient-facing systems. When should a conversational agent have a humanlike appearance or any physical embodiment? Would allowing for unconstrained speech and gesture input improve intervention effectiveness or reduce patient safety and privacy? When are nonverbal cues significant in designing a conversational agent? Does age, gender, socioeconomic status, literacy levels, or certain characteristics of target users mediate the effectiveness of VH interventions?

Finally, we found that the input and output of VH systems have evolved significantly over the last 2 decades, drawing on the most recent technological advancements. While systems in the 2000s extensively used desktops, kiosks, 2-dimensional graphics, and constrained text input [64], more recent systems were developed in 3D, were delivered in VR, sometimes used HMDs, and allowed for unconstrained input, such as speech, gestures, and facial emotions. Two key design trends were identified: (1) multimodal sensing of the user’s state using computer vision algorithms and ubiquitous computing technologies, ranging from upper-body gestures to heart rate (input), and (2) striving toward high fidelity, humanlike appearance and behavior of the VH to improve presence and immersion (output).

The limitations of this review should be noted and can be addressed by future studies. First, not all studies on VH in patient-facing systems were included in our work. This is because they did not present sufficient quantitative information, only reported usability metrics, or did not clarify whether their avatar technology was computer-controlled or human-controlled. Including additional studies and VH designs could allow reinforcing the results reported here or provide different results. Second, some of the studies included in our meta-analysis had relatively small sample sizes (<20 participants); thus, additional caution is recommended when generalizing these results. Third, there was moderate heterogeneity among trials in the meta-analysis (τ²=.12, I²=49.3%.). This can be attributed to the different health outcome measures, population types, and health objectives. For example, health objectives ranged from increasing physical activity and improving mood to improving social skills. When more studies become available with the same or similar health objectives, it would be worth updating this study with new results.

Although research on conversational agents began circa 2000, their design and capabilities have changed and diverged substantially as new technologies and sensors have emerged. This change is expected to continue. Future studies are suggested to consider the difference between different types of conversational agents when synthesizing or generalizing the agents. For example, does a physical appearance or nonverbal behavior increase the effectiveness of a conversational agent? In what kind of tasks? Furthermore, there is rich literature on behavior change and health behavior change theories. However, theoretical frameworks explicating how different features of VH work together in building (or disrupting) rapport with patients is lacking. As such models emerge, future studies will need to examine those relationships between model constructs with methods such as meta-analytic structural equation models.

VH are conversational agents with a humanlike physical appearance; autonomy in verbal and nonverbal behavior; and speech, gaze, or gesture interaction capabilities. In patient-facing systems, they can demonstrate listening and empathy, as well as tailor to various sociocultural backgrounds, languages, and literacy levels. We surveyed the existing literature on VH in patient-facing systems — from inception to December 2019. Of the 53 articles reviewed, a meta-analysis of 26 studies with more than 1400 participants showed that VH interventions significantly improve health outcomes compared with other traditional intervention methods. But whether a physical embodiment is crucial for a conversational agent to significantly improve health-related outcomes remains to be explored, as does any effect of the VH’s physical appearance, type of voice, or quality of movements.

Although not yet comparable to computer-animated films or high-end video games, the appearance and behavior of VH in health care are increasingly becoming sophisticated, with studies finding that users prefer more humanlike VHs in health care [124,125]. Elsewhere, studies continue to report the possibility of unintended negative user reactions (ie, the uncanny valley effect) when interacting with VH, owing to a mismatch in the levels of realism, either physical, behavioral, or both [23,24,126]. However, the literature on VH in patient-facing systems has not yet examined whether the uncanny valley effect affects patient perception and, in turn, health outcomes.