We present a taxonomy of VHs in mental health VR (Figure 1), with categorization of their features into 3 main aspects: role, agency, and interaction type. These features were selected because they represent the critical dimensions influencing how VHs are designed, implemented, and experienced in mental health VR. Role defines the primary function of the VH within the virtual environment, agency determines the level of control and autonomy the VH exhibits, and interaction type describes the nature and depth of engagement between the VH and the participant. Together, these features provide a framework for understanding the diverse ways VHs are used in mental health research, enabling researchers to systematically evaluate their design and impact.

This framework was inductively derived from the included studies, and it was also supported by prior classification approaches in VR and virtual character research. Existing taxonomies highlight comparable dimensions—distinguishing characters according to their functional role [29,30], level of autonomy or agency [19,29,31,32], and communicative capacity or interaction features [19,30]. By aligning with these established design variables while adapting them to the specific context of mental health VR, our taxonomy provides both empirical grounding and theoretical consistency with prior frameworks.

The inclusion and exclusion criteria for eligible studies are summarized in Textbox 1. In brief, studies were required to (1) use VHs within an immersive VR setting; (2) involve assessment or intervention related to mental health conditions; and (3) include empirical data from human participants. Studies were excluded if they were not related to mental health, did not involve a VR-based VH, or were not available as peer-reviewed full-text articles in English.

The literature search was performed in PubMed, Scopus, MEDLINE (Ovid), PsycINFO (Ovid), and Web of Science (Clarivate), with title and abstract searches for keyword combinations involving (1) VR, (2) VH, and (3) mental health. The keyword query was structured as follows: [(virtual reality OR immersive virtual reality OR VR) AND (virtual human OR virtual character OR virtual agent OR avatar OR humanoid) AND ((assessment OR treatment OR therapy OR mental health) OR (mood disorders OR depress* OR bipolar OR mania OR paranoia OR psychosis OR psychotic OR schizophren* OR schizotyp* OR delus* OR hallucinat* OR phobias OR obsessive compulsive disorder OR OCD OR anxiety OR post traumatic stress disorder OR PTSD OR trauma OR anorexia nervosa OR bulimia nervosa OR eating disorders OR binge eating OR insomnia OR sleep OR nightmares OR circadian OR panic OR substance OR abuse OR cannabis OR tobacco OR alcohol OR amphetamine OR hallucinogens OR heroin))]. Details of the searches for each database in the corresponding platforms can be found in Multimedia Appendix 1. A completed PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist [50] is presented in Checklist 1.

Following database searches, we identified 2596 records (Scopus, n=1232; Web of Science, n=712; PubMed, n=233; APA PsycINFO, n=247; MEDLINE, n=172) between 1997 and 2024 (final search: November 20, 2024). After removing 1085 duplicates (11 manually and 1074 via Covidence), 1511 unique records remained. Title and abstract screening excluded 1280 records that did not involve a VH character, did not use immersive VR, were unrelated to mental health, were nonhuman studies, or were non-English studies. The remaining 231 full-text articles were then assessed in detail, with the reasons for exclusion including the following: not immersive VR (n=73), not related to mental health (n=31), not full paper (eg, conference abstract) (n=21), no empirical data (n=19), no VH (n=11), and case study with only 1 participant (n=3).

Figure 2 (PRISMA diagram [50]) summarizes the search and screening results. In total, 79 studies were included in the review: 73 that met all eligibility criteria at full-text screening and 6 identified through citation chaining. We have provided a list of studies that primarily used VHs as active social interaction partners (Table 1), virtual crowds (Table 2), and virtual bodies (Table 3), along with a brief summary for each.

VHs were used in the following research areas in mental health: social anxiety (n=20), eating disorders (n=18), psychosis (n=17), substance abuse (n=5), phobias (n=4), depression (n=3), autism (n=3), PTSD (n=2), dissociative disorders (n=1), and others without specific mental health conditions (n=6).

We categorized the studies based on our taxonomy. Studies with overlapping roles were assigned to their primary role or the one with the highest interactivity. The roles of VHs in the included studies could be categorized into active social interaction partners (n=40), virtual crowds (n=16), and virtual bodies (n=23).

Sixteen studies used virtual crowds. Six of them were related to anxiety. In 1 of the earliest studies, Slater et al [3] compared public speaking to an attentive versus a disruptive virtual audience, finding that crowds could evoke anxiety comparable to real life, with better performance when the audience was attentive. Similar findings were reported by Takac et al [87], who suggested that the distress created in a virtual public speaking scenario could have lingering effects after the VR session. Virtual crowds were also used to populate a virtual office, finding that watching other VHs working was effective in triggering work-related stress [49] or creating peer effects to make people work more efficiently [14]. Brinkman et al [48] explored the effect of the size and ethnicity of virtual crowds in a bar scene, showing that VR exposure to an increased density and proportion of VHs with other ethnicities could induce stronger distress. Additionally, Shiban et al [84] incorporated air blasts and the sound of a female character screaming when participants approached a virtual crowd to study how individuals develop social fears and how these fears can be diminished or overcome through repeated exposure.

Six studies used a virtual crowd for research on psychosis. Early tests examined people’s appraisals of neutral characters in a library scene, demonstrating that virtual crowds elicited persecutory thoughts for insights into psychosis [81,82]. Later research employed daily VR scenarios populated with VHs, such as in an underground train or school canteen, to understand factors associated with paranoid thinking. The findings suggested that paranoia could be predicted by anxiety, worry, perceptual anomalies, and cognitive inflexibility [15], and was influenced by individuals’ self-confidence [86]. For adolescents, interpersonal threat or hostility heightened state paranoia [89]. Further, Wei et al [88,111] found that VH facial design influenced both paranoia and visual attention.

Mountford et al [85] used a virtual crowd in a virtual bus ride to study eating disorders. Participants were tested for their body image satisfaction, and the results showed that exposure to the virtual scenario with VHs did not necessarily change people’s current feelings or perceptions about their body image. Cho et al [83] designed a VR bar to study alcohol cravings and found that the introduction of drinking characters led to a stronger drinking desire compared to simply displaying the alcohol in the scene. Similarly, Rovira et al [90] found that VR exposure to smoking-cue VHs elicited stronger cigarette cravings than neutral environments without VHs. Rizzo et al [39] used virtual soldiers and civilians for military training and found a reduction in PTSD severity in 80% of the study completers.

VHs were used as body representations in 20 studies, with 15 focusing on eating disorders. Riva and Melis [1] conducted the first study in the literature on body image representation using VR. Participants selected virtual figures representing their current and ideal body size. Perpiñá et al [91] introduced a feature allowing patients with eating disorders to adjust the size of a virtual body to reflect their body image. Both showed VR’s feasibility for assessing body image disturbances, with results comparable to paper-based measures but showing greater engagement [18]. While these early studies did not induce body ownership, later studies often aimed to do so.

Studies employing the body ownership illusion paradigm found that embodying a larger virtual figure led participants to check their virtual body more frequently [94,95] and experience a higher degree of self-dissatisfaction and anxiety [40,96,98,101,108]. In contrast, Schroeder et al [110] found that body dissatisfaction was not necessarily affected by the weight of the avatar, although people with higher body dissatisfaction exhibited more body-checking behaviors in weight-related areas.

Ferrer-Garcia et al [93] created a paradigm to reduce anxiety related to body image in female students by making them return to the normal-size virtual figure after experiencing a larger body. A similar method was used to alleviate body image distortion in both healthy-weight participants and those with obesity [99]. Wolf et al [103] compared the body weight perceptions of 2 groups of females: one that embodied a VH and another that observed the VH as another person’s avatar. The self-embodiment group underestimated the body weight of the VH compared to those who rated the VH as another’s avatar. Furthermore, Ascione et al [109] conducted a body-related attention bias modification task to reduce excessive visual focus on the weight-related areas of a self-represented VR body, which significantly decreased body dissatisfaction.

Three other studies used virtual bodies for psychosis research. Spanlang et al [100] examined whether self-fragmentation in VR, where participants embodied a VH while retaining physical presence in the real world, impacted the biological responses in patients with schizophrenia. Yamamoto and Nakao [105] found that perceiving a fake body as one’s own reduced body ownership in depersonalization. Gorisse et al [102] used a virtual body double resembling each participant for a social VR task. They concluded that observing a body double engaging in VR social interactions could reduce persecutory thoughts during the task.

Additionally, van Gelder et al [104] demonstrated that alternating convicted offenders between their current and aged future selves reduced self-defeating behaviors. Similarly, embodying an older body could increase social motivation among young adults [107]. Besides, Burin et al [106] presented participants with a moving character from first-person and third-person perspectives. They concluded that a body illusion from the first-person perspective created stronger physiological activation to help people practice stress coping. Aymerich-Franch et al [92] assessed the impact of the facial similarity of a VH on participants, finding that embodying a figure with a dissimilar face reduced anxiety during VR presentations.

Out of the 79 studies, 74 involved the use of VAs. Nine of them adopted a semiautonomous agent design, where facilitators manually triggered an audio effect [3,61], adjusted VH voices [75], or played prerecorded animations [4,55] to customize characters’ behaviors in real time.

Regarding the interaction types of VAs, approximately half (39/74, 53%) engaged in explicit interactions with participants, primarily serving as active interaction partners. In 10 studies, VAs exhibited implicit interactions (eg, occasionally looking at participants and then looking away). Passive interactions were observed in 26 studies where VHs were used as virtual crowds and virtual bodies.

In the reviewed studies, the term “avatar” was often used to describe any type of VH, even when they did not represent the participant’s self. Only 5 studies used VHs as participant avatars. For example, Aymerich-Franch et al [92] used an avatar with varying levels of visual similarity to the participant’s face to study the self-embodiment experience. Additionally, 12 studies combined the use of agents and avatars to create perspective changes through body-swapping experiences in social interactions [11,13,104] or to explore the effects of vicarious agency (the perception of control over the VH) [102]. Kothgassner et al [45,64] further examined the effect of perceived agency by comparing participants’ VR experiences with VAs and avatars. The results indicated that avatars could elicit stronger emotional responses compared to VAs.

Of the 79 studies reviewed, 44 examined the effects of manipulating specific characteristics of VHs. These manipulations aimed to assess how physical and behavioral attributes of VHs influence participants’ perceptions and behavioral responses. The manipulated features included their visual appearance [48,52,99] (eg, gender, ethnicity, and body size), interaction behavior [7,76,87] (eg, interpersonal distance, eye gaze behavior patterns, and responsiveness toward participants), perceived agency [45,64] (VA or avatar), and emotions [52,75] (eg, emotional facial expressions) or personality [58]. While most of the traits were nonverbal behaviors, 4 studies explored the effects of verbal behaviors, including dialog content [13], conversation attitude (eg, positive vs neutral dialog) [56,57], and the voice of VHs [75]. The most frequently studied trait was body size (n=13), while traits, such as sense of affirmation [37] and personality [58], received limited investigation.

In addition to individual traits, 2 studies examined the effects of crowd size and density as independent variables in virtual crowd settings [48,87]. These studies explored how variations in the number and arrangement of virtual characters influence participants’ sense of presence and anxiety in crowded environments.

This paper reviewed VR studies that included VHs in scenarios to understand and treat mental health difficulties. We examined 79 studies and categorized them based on the primary role of the VHs: active social interaction partners, virtual crowds, and virtual bodies. We focused on the use of agency, types of interactions, and characteristics of the VHs. VHs have served diverse purposes and demonstrated the ability to create psychological, behavioral, and physiological influences. However, few studies have provided comprehensive descriptions of VHs’ visual appearance or behavioral capabilities. Without such details, it becomes challenging to assess their actual impact, and replicability is reduced. The potential of each VH role requires further exploration to address these gaps and the associated challenges.

Active social interaction partners interact explicitly with participants. A persistent challenge in this context has been enhancing the realism and responsiveness of these interactions, particularly during close-up engagements [19,112,113]. Improving both visual fidelity and behavioral plausibility is essential for achieving lifelike interactions [30,114,115]. Consistent design of both visual and behavioral elements is needed to avoid the uncanny valley, where characters appear almost human but evoke discomfort due to subtle imperfections, as also seen in VR interactions [116,117]. Additionally, it is important to consider VH behavioral plausibility and consistency across multiple behavioral channels, including both verbal and nonverbal communication (eg, body movement, facial expressions, voice, and dialog content). For example, Choudhary et al [118] demonstrated that mismatches between facial expressions and vocal tone (eg, a happy face with an unhappy voice) negatively impacted participants’ immersion and engagement, highlighting the importance of aligning these communication modes.

Regarding responsiveness and interactivity, current feedback in mental health VR often relies on prerecorded actions triggered by participant inputs, such as user interface selection, position, or response time [49,60,65]. Future applications could integrate machine learning algorithms, such as reinforcement learning, to create autonomous VHs that learn from experience [119]. These VHs could adjust parameters, such as the interpersonal distance, to maximize attention and keep participants engaged. Algorithms could detect interaction dynamics (eg, participant attitudes and responsiveness) and adjust VH behaviors in real-time to achieve specific therapeutic goals, particularly in applications encouraging positive behaviors. Another area to explore is the real-time monitoring of physiological responses to create adaptive interactions based on participants’ affective state. A few of the studies reviewed used biometric data (eg, salivary alpha-amylase, heart rate, and galvanic skin) to evaluate physiological responses [52,100,106], but none leveraged these data to enhance interaction fidelity. Real-time estimates of emotional arousal and valence could enable VHs to respond meaningfully to participants’ emotional changes [120].

A virtual coach represents a specific type of active social interaction partner, designed to guide and motivate users toward positive behavioral changes while facilitating the automation of therapy delivery. Focus group feedback indicated that users valued the convenience, accessibility, and anonymity provided by a virtual coach, but they also expressed concerns about the coach feeling unrelatable and difficult to engage with [121]. The development of a VR coach should incorporate knowledge and feedback from domain experts and targeted user groups to ensure it is tailored to the specific age range and mental health conditions it aims to address [122,123]. Generally, increasing the engagement and friendliness of communication from virtual coaches is likely to improve their effectiveness and the outcomes of VR therapy [68].

In virtual crowd studies, it is standard practice to use identical designs and behaviors for characters that share common objectives to achieve a balance between realism and computational efficiency [124,125]. However, visual and behavioral repetition patterns in a group of VHs are very noticeable and decrease fidelity. Introducing variations in appearance and movement is important to diversify crowds and avoid these patterns [126,127]. Visual diversity can be achieved, for example, by randomizing height, body shape, hair color, and clothing. Behavioral diversity can be enhanced by blending animations, using motion capture data from multiple individuals, and implementing procedural animations with a certain degree of randomization to generate varied movements in real-time based on environmental and interaction dynamics.

There are additional considerations. When virtual crowds have autonomy to move around a space, each one should navigate naturally, avoiding collisions with static and dynamic obstacles. For example, Trivedi and Mousas [128] implemented a crowd behavior system for a VR street scene, where each VA had navigation scripts with steering and pathfinding capabilities to avoid collisions. Each agent also had a customizable target location, walking speed, and animation (eg, head direction and gaze pattern). Other advanced virtual crowd control systems generate real-time behaviors driven by user input and the dynamics of other virtual entities within the crowd [129]. These designs could be applied in mental health applications to improve crowd interactions and better understand responses to busy environments.

Studies involving virtual bodies predominantly helped individuals reassess their body image biases or gain new perspectives, provided by the body ownership illusion. These use cases require improved methods to enable the sense of embodiment. Existing methods like visual-motor synchrony [97] and visual-tactile synchrony [96,99] are designed to align visual feedback with participants’ movements or tactile sensations to enhance the illusion of ownership over the virtual body. While effective in initiating this illusion, sustaining or manipulating it throughout the VR experience remains challenging. Mismatched sensory feedback or slight discrepancies in motion can disrupt the sense of embodiment [130,131]. Integrating other sensory modalities, such as auditory feedback (eg, footsteps and voice modulation) or olfactory feedback (eg, simulating a scent experience based on the participant’s location), could enhance the body ownership illusion. Combined multisensory feedback has been shown to strengthen and prolong the influence on embodiment, affect, and behaviors [132,133].

Customizing virtual bodies to closely resemble participants can enhance the sense of embodiment. Aymerich-Franch et al [92] found that embodying a virtual body with a similar face increased self-association and anxiety during public speaking tasks compared to dissimilar faces. This provides evidence for the importance of avatar personalization and its potential to modulate emotional responses. Customization should consider not only physical appearance but also factors like ethnic background [134] and voice [135]. These factors can significantly contribute to the body ownership illusion.

The integration of VHs in mental health VR raises potential ethical and methodological challenges. A key concern is the lack of standardization in the technical implementation of character creation and animation. Many implementations rely on proprietary or nonstandardized software, which complicates replication and cross-study comparison. When such tools become inaccessible or difficult to adopt, scaling the experiments or building cumulative evidence becomes challenging. Moving toward open-source platforms or developing shared technical standards would enhance reproducibility and research efficiency. Encouragingly, the computer science community has started to introduce open-source 3D character libraries [26-28], some of which have already been used in mental health research [11,13]. This provides early examples of community-driven resources that could serve as a foundation for broader standardization.

Equally important is the transparency of reporting with regard to VHs. Many current studies provide limited descriptions of VH characteristics, and some even omit a general description of VR scenes. Such omissions reduce study transparency and limit replicability for continuous research. General ethical guidelines from the UK National Institute for Health and Care Excellence (NICE) [136] emphasize clear reporting of system design and capabilities for therapeutic applications. The CONSORT-EHEALTH extension recommends that digital health interventions specify intervention components, delivery modes, tailoring, and technical features [137]. More specifically, RATE-XR (reporting for the early-phase clinical evaluation of applications using extended reality) provides a reporting framework for early-phase clinical extended reality applications, with a checklist covering application characteristics, hardware, software, and development details [138]. For more tailored standards, VH-specific reporting practices, ideally supported by visual and technical documentation (eg, screenshots, videos, and model specifications), are needed to improve reproducibility and comparability across disciplines.

Moreover, data privacy and participant safeguards also require careful consideration. VH-mediated applications can sometimes collect sensitive physiological, behavioral, and conversational data [13,48,139], raising risks related to identity inference and emotional profiling, where user behavior data can sometimes be used to infer sensitive personal traits [140]. In addition, embodiment-based interactions can sometimes provoke unintended psychological effects, such as distress or overidentification with the VH [141,142]. To mitigate these risks, researchers should implement robust informed consent procedures and ensure secure data handling. Safeguards, such as participant prescreening, clinician oversight, built-in safety controls (eg, pause or exit functions in VR software), and structured debriefing protocols, are essential to protect participants.

This systematic review has several limitations. First, our search strategy may not have captured all relevant VR mental health studies involving VHs. We used 5 databases and conducted searches limited to titles and abstracts with fixed keyword combinations, excluding studies that did not simultaneously mention VR, VHs, and mental health. To maintain a consistent search strategy across platforms, we did not use controlled vocabulary, such as Medical Subject Headings (MeSH) in PubMed, which may have further reduced recall. To mitigate this, we conducted citation chaining and identified 6 additional eligible studies not captured in the original search. However, the omission may have still remained, and it may have narrowed the evidence base. Future updates should refine the search strategy by incorporating broader synonyms and controlled vocabulary, and including gray literature sources.

Second, there are limitations in our categorization of VH roles and characteristics. Existing taxonomies used to classify virtual characters and their behaviors [30,143] are at least a decade old and do not specifically focus on VR studies in mental health. We developed our own classification based on the usage of VHs in the reviewed studies. This framework is tailored to mental health research, and it may require adaptation for application in other research areas. Ongoing efforts to develop and validate updated, domain-general taxonomies would improve consistency and comparability across studies.

Third, most included studies were pilot, feasibility, or exploratory in nature, often with small sample sizes and limited follow-up. While we included only empirical studies with more than one participant to enhance rigor, the predominance of early-stage designs may have constrained overall study quality and generalizability. Future research could prioritize more robust designs, such as randomized controlled trials or longitudinal studies with quality assessments, to generate stronger and more interpretable evidence.

Fourth, regarding result analysis, we opted for a narrative review approach instead of a quantitative evaluation to measure the quality and effectiveness of VHs across studies. This choice was made due to the variability in study objectives (ranging from validation and assessment to clinical treatments) and the diversity in participant characteristics. In addition, many studies lacked sufficient descriptions of VH attributes and consistent reporting on VR systems and software, which constrained systematic comparison and aggregation. As the field progresses and reporting becomes more standardized, meta-analyses and formal quality appraisals will become feasible.

VHs can be used in many different ways within immersive VR mental health research and have served primarily as active interaction partners, virtual crowds, and virtual bodies. Across the 79 studies reviewed, most VHs were computer-controlled agents, with perceived agency shaping emotional and behavioral responses. Research to date has predominantly examined body size and gender, with limited attention to emotional expression or personality traits. Many studies provided insufficient technical and visual details on VH design, limiting reproducibility and cross-study comparison. Addressing these gaps is crucial for advancing the evidence base and understanding how specific VH characteristics influence mental health outcomes in VR. To our knowledge, this review provides the first synthesis of the use of VHs in VR mental health, offering a potential foundation for future VR programming of VHs and future studies on their implementation.