In this review, the definition of digital human is equivalent to the definition of virtual human. As described by Traum [19], a virtual human is an “artificial agent that includes both a visual body with a humanlike appearance and range of observable behaviors, and a cognitive component that can make decisions and control the behaviors to engage in human-like activities.” Although this definition is comprehensive, there are still some ambiguities, such as the judgment of humanlike appearance and humanlike activities. To provide a clearer set of criteria to assist us in filtering articles, we have defined a digital human as an agent with 3 criteria inspired by the 10 traits suggested by Burden and Savin-Baden [9]. A digital human (1) visually possesses realistic appearance characteristics of a human, including facial features and body proportions; (2) is capable of performing nonverbal behaviors, including body movements and facial expressions; and (3) must engage in bidirectional communication, understanding, and responding to verbal and nonverbal cues from users.

If the virtual collocutor described in the reviewed article meets all of the criteria, then we consider it to be a digital human.

Support for depression encompasses a holistic approach that integrates both direct and indirect methods to aid individuals in managing and overcoming the condition. This comprehensive support system is essential for addressing the multifaceted nature of depression, which affects individuals emotionally, physically, and socially [3]. Following the framework established by Cuijpers [20], we distinguish between direct and indirect support based on their primary target. Indirect support focuses on problems related to depression—such as social isolation, lifestyle factors, or caregiver burden—where interventions address associated factors rather than depression as the primary clinical target. In contrast, direct support explicitly targets depression assessment (eg, administering Patient Health Questionnaire-9 [PHQ-9], conducting diagnostic interviews based on Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition [DSM-5] criteria) or depression-specific therapeutic interventions (eg, cognitive behavioral therapy [CBT] for negative thought patterns and compassion-focused therapy [CFT] for self-criticism). In this paper, we narrow the scope to direct support provided to individuals with depression, as this aligns with our RQs that specifically examine how digital humans function within depression assessment and intervention protocols. This means that the primary contribution of selected papers is to provide assistance to individuals undergoing assessment for depression and receiving treatment for it.

The applied protocol in assessment can be mainly divided into 3 categories—questionnaire, interview, and computer-based interactive task.

The first category is questionnaires, accounting for 3 of the studies [21,23,33]. In this category, the digital human will ask users the questionnaire questions one by one and collect their responses. Only answers that correspond to existing options will be collected; otherwise, the digital human will continue asking the user until obtaining a usable answer. The questionnaires that were used include PHQ-9 [43] and Center for Epidemiologic Studies Depression Scale - Visual Analog Scale - Very Short version (CES-D-VAS-VS) [44]. PHQ-9 is a brief self-report tool that consists of 9 items, each of which is scored on a scale from 0 (not at all) to 3 (nearly every day), which are directly derived from the diagnostic criteria for major depressive disorder in the DSM. The CES-D-VAS-VS is an adaptation of the traditional Center for Epidemiologic Studies - Depression Scale, which is a 20-item questionnaire used to measure depressive symptoms in the general population. The CES-D-VAS-VS incorporates a visual analog scale, enhancing its sensitivity by allowing patients to mark their symptom severity along a continuum.

The second category is assessments through interviews, accounting for 5 of the studies [26,28,29,31,39]. As an essential tool in depression assessment, interviews are commonly used in clinical settings where a nuanced understanding of the patient’s condition is crucial [45]. In the 5 studies that used interviews for depression assessment, 2 [26,28] focused on interviewing about negative topics, such as war and loneliness. These topics were chosen because they significantly impact vocal, visual, and verbal features, which are critical in detecting depression. The study by DeVault et al [29] based its interview content on observations from face-to-face interviews conducted in a clinical setting. This approach aimed to mimic real-life interactions and assess how patients express their symptoms naturally. Wu et al [31], through collaboration with clinicians, designed 93 questions about depression, anxiety, hypomania, and family relationships. This series of questions was specifically designed to vary depending on positive, negative, and neutral emotions. The study by Philip et al [39] structured its interview around the DSM-5 [46] criteria for major depressive disorder. This method ensured that the interviews were comprehensive and aligned with established diagnostic standards, facilitating a more systematic approach to identifying depressive symptoms based on the latest psychiatric guidelines.

A computer-based interactive task was only used in 1 study [22]. In this study, researchers and mental health experts collaboratively designed 4 different types of tasks—mimicking, dyadic interaction, digital treatment, and psychometric. The design of these tasks aimed to be sufficiently engaging to prompt use without the need for physician guidance, executable on mobile devices, and effective at eliciting behavioral traits relevant to digitally assess mental health.

There are a total of 13 (65%) studies focusing on the intervention phase. The therapeutic methods involved include CBT [23,25,27,33,37,40,47], psychoeducation [34,36,38], CFT [24,35], and avatar therapy [30,32].

CBT [48], as one of the most common therapy approaches in studies, aimed at teaching individuals how to recognize negative patterns of thought, evaluate their validity, and replace them with healthier ways of thinking. Among all the studies that applied CBT, they mainly focused on cognitive restructuring [49] and behavioral activation techniques [50]. In cognitive restructuring, participants need to recognize negative thoughts related to a recent event and are then encouraged to contemplate a more positive interpretation, and behavioral activation is a CBT technique that motivates participants to undertake avoided activities and engage in tasks that bring pleasure and accomplishment. Luerssen and Hawke [25] applied low-intensity CBT [51] in their study. Compared with traditional CBT, low-intensity CBT involves simpler interventions that can be delivered by practitioners who are not necessarily clinical psychologists and is more standardized, using general strategies that apply broadly to everyone with similar symptoms, while CBT is highly personalized and involves developing a specific therapeutic strategy for each patient’s unique problems and needs.

Psychoeducation is a fundamental component of depression intervention that involves educating individuals about depression as a disorder, including its symptoms, causes, and treatment options [52]. In the studies by Shamekhi et al [34,38], the content of psychoeducation primarily focuses on the management of nutrition, physical activity, pain, stress, sleep, and depression and includes guiding patients through practice sessions, such as meditation and yoga. Meanwhile, in the study by Bickmore et al [36], the educational content focused on the postdischarge self-care regimen, including medications, follow-up appointments, exercise and diet regimens, and pending laboratory tests.

In the remaining 4 studies, 2 [24,35] used CFT [42], a psychological approach designed to promote mental and emotional healing by encouraging individuals to develop compassion for themselves and others. The other 2 studies [30,32] used avatar therapy, where therapists interact with clients using a computer-generated avatar. Ashrafi et al [30] conducted a computer-assisted avatar-based treatment for dysfunctional beliefs (CAT-DB), which uses an avatar to help patients engage in dialogue with their dysfunctional beliefs and confront them.

In studies applying digital humans to assessment, digital humans have been proven to be as effective in conducting interviews as showing real human videos [26,28] and filling out self- report questionnaires [21]. In the studies by DeVault et al [29] and Philip et al [39], data collected after interviews with digital humans were analyzed, and they concluded that these data have the potential to support automatic depression assessment based on dialogue systems involving digital humans and humans. Furthermore, in the experiment by Egede et al [22], it was found that participants exhibited greater body movements and more intense facial expressions under the guidance of a digital human (compared with text-only mode), which also supports the feasibility of using digital humans for automatic depression assessment.

Many experimental results that apply digital humans in interventions have highlighted the potential for digital humans to establish positive relationships with participants. This includes factors such as positive responses to participants’ emotional expressions, customizable appearances, nonjudgmental expressions, and good patience. This personal relationship can also lead to better acceptance and attitudes toward digital humans from participants. Several articles reported that digital humans made a positive impression on participants. In the experiment by Bickmore et al [36], 76% of participants reported that their experience talking with a digital human was better than their interactions with health providers, and they were more inclined to choose digital humans as their channel for receiving psychoeducation in the future. Additionally, several articles indicated a positive impact of digital humans on therapeutic outcomes. For example, in the experiment by Kocur et al [32], they compared the results of participants receiving treatment as usual (TAU) with those receiving CAT-DB+TAU and found significant differences in the reduction of depressive symptoms in the group who received CAT-DB+TAU.

The design of digital humans across various studies showcases a diverse range of models, display devices, input or output modalities, and back-end intelligence, reflecting the adaptability of virtual characters to different applications. Based on the description of digital humans in the studies, we identified that the digital human designs in 3 studies [23,26,34] appeared repeatedly. We chose to retain the literature with more detailed descriptions of the digital human designs. After eliminating these duplicates, 17 unique digital human designs were included in the results.

In terms of body visibility, there are 7 full-body, 8 upper-body (head to chest), and 2 head-only digital humans (Figure 2). After analyzing the use cases for each digital human, we found that on mobile devices (such as phones and tablets), half-body is a more common setting [31,37], as this aligns better with how we typically see human bodies on these devices. The choice of full body is usually driven by device or scenario requirements. For example, digital humans in virtual reality (VR) all use full-body representations [24,35,40], and in the study by Stratou et al [53], they displayed a full-body digital human in larger screens to simulate the feeling of conversing with real people in reality. As mentioned by Bresó et al [27], they chose head-only because focusing on facial expressions makes it easier to convey different emotions and gives the digital human a higher degree of realism, which is a key issue in their study design. Regarding style design, the appearance of 2 digital humans leaned more toward a cartoon-like style [25,28]. Luerssen and Hawke [25] mentioned that their choice of cartoon-like style was to avoid the uncanny valley effect, while the rest of the digital humans adopted a more realistic style.

The majority of digital humans (n=15, 88.2%) support natural language input, with 2 studies [36,38] relying on interface touch input. Correspondingly, all digital humans support speech synthesis and audio output, but only 5 digital humans are capable of delivering empathy narration [22,27,29,30,35]. All digital humans exhibit lip movements while speaking, along with other bodily actions, such as blinking [28] and nodding [29]. Due to their presence in a 3D environment, the digital humans in the studies by Baghaei et al [24], Halim et al [35], and Wu et al [31] are capable of a wider range of actions, including walking, crying, showing anger, and more.

This scoping review identifies 4 types of back-end intelligence used in the development of digital humans—Scripted Systems, Rule-Based Systems, Wizard of Oz, and Recorded Playback. Each approach offers unique strengths and applications, contributing to the diverse landscape of digital human technology.

Scripted systems operate based on predetermined scripts that define the digital human’s behavior and responses in a linear, predictable manner. This method is highly effective in scenarios where interactions are straightforward and the range of possible user inputs is limited. Among the 17 digital humans, more than half (n=10, 58.8%) used a scripted back end to control their behavior. For the digital human who used a scripted back end, the scenarios they face have standard dialogue processes, such as interviews with predefined topics [26,31], and tasks with specific instructions [22].

Rule-based systems are characterized by a set of predefined rules that guide the digital human’s decision-making processes. These systems offer greater flexibility compared with scripted systems, as they can dynamically adapt to a wider range of user inputs through conditional logic. Furthermore, 4 digital humans use a rule-based dialogue system, with 3 digital humans acting as counselors in the studies [25,27,37], and 1 as an interviewer [21].

The Wizard of Oz [54] technique involves a human operator who controls the digital human’s actions and responses in real-time, unbeknownst to the user. This method is commonly used in experimental and prototyping phases, allowing researchers to simulate the capabilities of a fully autonomous digital human before the underlying technology is fully developed. The digital humans in the experiments conducted by DeVault et al [29], Kocur et al [32], and Hidding et al [40] are directly controlled by professionals (certified counselors) to ensure participant safety and provide an experience closer to interacting with a real person.

Notably, in the experiment by Halim et al [35], the behavior of the digital human is derived from the participants’ actions, a method called Recorded Playback. In this experiment, participants are required to show compassion to the digital human in the first stage, and the language and actions of the participants are recorded by the system. In the second stage, these recorded behaviors are performed by the digital human.

The aim of this scoping review was to identify the existing evidence on using digital humans to assist in assessment for depression and delivering interventions, and to unfold the design choices of these digital humans. We have found that many studies designed digital humans’ purposes to be similar to certain roles in existing services. For example, a counselor conducting assessment through interviews [31], a therapist providing CBT session [27], and a nurse giving necessary information to patients [36]. The results of these studies can be considered preliminary affirmations of the effectiveness of digital humans in these roles.

Advantages of digital humans: Compared with real humans, digital humans have significant advantages in accessibility and availability, as they are essentially software programs. This means they can operate around the clock without the constraints of human limitations such as fatigue, working hours, or geographic location. This 24×7 availability ensures that support can be provided to users whenever they need it, which is particularly beneficial in emergency situations or for individuals in different time zones. Additionally, applying digital humans in the field of psychotherapy has a special advantage—the low social stigma caused by interaction. In selected studies, 2 articles [21,36] mentioned that in experimental interviews comparing interactions with digital humans and health providers (therapists or nurses), several participants felt more relaxed and safe communicating with the digital human. This advantage has been proven by the research by Lucas et al [55], which demonstrated that participants were more willing to disclose emotionally sensitive information, including expressing sadness more intensely, when they were interviewed by a digital human and believed the interviewer was a computer rather than a human. Another work by Loveys et al [56] indicates that the emotional expressiveness of digital humans, such as using an emotional voice, can enhance participants’ comfort and emotional responses during interactions. These studies suggest that digital humans can create an environment where individuals feel more comfortable and less anxious when disclosing. Furthermore, the consistency and unbiased nature of digital humans ensure that all users receive the same level of care and attention, free from the potential biases or variability that can come with human providers. This standardization can be particularly important in ensuring equitable access to high-quality support services.

Designed for real-life scenarios: In the design of these digital humans, a key design choice has been to ensure that the digital human’s appearance and behavior align with the local cultural context. For example, Takemoto et al [26] specifically mentioned that they designed the digital human’s appearance based on the local population’s characteristics. In the study by Wolters et al [23], besides the appearance, the digital human was given a Scottish accent to resonate more effectively with participants from that region. Based on previous research, such culturally sensitive design decisions can significantly enhance participants’ comfort and engagement, potentially leading to more effective therapeutic outcomes [57,58]. Moreover, the input modality is another vital consideration. While speech is often the natural and preferred input method of digital humans, some studies, particularly those conducted in noisy environments like hospitals [36], have opted for touch as the primary mode of interaction. This choice underscores the importance of usability in designing digital humans, emphasizing that how users interact with these entities must be carefully tailored to the specific context in which they are deployed.

Leveraging the advantages of digital humans, we believe that their potential can be extended to a wider range of roles to assist in depression assessment and intervention, especially in helping clients reduce social stigma. For instance, in face-to-face therapy, some clients may bring a support person to the session to help alleviate their anxiety. A digital human can replicate this support function by providing a nonjudgmental and consistent presence, helping clients feel more comfortable and less anxious during their sessions. Another promising role for digital humans is that of a peer specialist. Peer support interventions have been shown to be effective in treating depression, as evidenced by numerous studies [59]. Peer specialists share their own experiences to help clients feel understood and to encourage them to disclose more during the intervention. A digital peer specialist can fulfill this role by simulating these supportive interactions, providing empathetic responses, and sharing relatable experiences that resonate with clients.

Digital human design choices reflect careful consideration of technological constraints, therapeutic goals, and user contexts. The analysis reveals systematic patterns in how researchers approach fundamental design decisions, from visual appearance to underlying intelligence architectures, with each choice serving specific therapeutic and technical purposes.

Body visibility decisions appear strategically aligned with platform capabilities and therapeutic objectives. Full-body representations dominate VR applications where spatial presence is paramount, while upper-body presentations are preferred for mobile devices where screen space is limited. Head-only implementations are chosen specifically when facial expression fidelity is critical for conveying emotions or avoiding uncanny valley effects [27,32]. This distribution suggests that successful digital human design requires careful consideration of platform affordances and their alignment with therapeutic goals. The predominance of realistic over cartoon-like styling (15 of 17 designs) suggests a preference for human-like appearance that supports therapeutic credibility and user engagement. However, some researchers deliberately chose cartoon styling to avoid uncanny valley effects while maintaining user comfort, indicating that appearance realism must be balanced against potential negative user reactions. Cultural sensitivity emerges as a critical design consideration, with researchers adapting appearance characteristics, accents, and interaction styles to match local population demographics and cultural expectations. These adaptations demonstrate awareness that effective therapeutic relationships require cultural alignment and familiarity.

Interaction modality design reveals sophisticated approaches to creating natural, accessible interfaces. The overwhelming preference for speech-based input (88.2% of implementations) reflects both user expectations for natural conversation and the technological maturity of speech recognition systems. Hence, the conversation ability is one of the main factors that affect closeness perceptions on digital humans [60]. The selective use of touch input in specific contexts, particularly noisy hospital environments, demonstrates adaptive design thinking that prioritizes usability over technological sophistication. This flexibility suggests that effective digital human systems must accommodate diverse deployment contexts and user needs.

Output modality design shows varying levels of sophistication in multimodal communication. While all systems provide speech synthesis and basic behavioral animations, the implementation of empathy narration in only 5 systems suggests that this remains a challenging but valuable design goal. The expanded behavioral repertoires available in 3D environments, including walking, crying, and showing anger, highlight how platform capabilities directly influence design possibilities and therapeutic potential. These variations indicate that output sophistication should align with both technical capabilities and therapeutic requirements.

The distribution of back-end intelligence approaches reveals important insights about current technological capabilities and design priorities. Scripted systems (10/17, 58.8%) dominate in contexts requiring predictable, standardized interactions, particularly for assessment protocols and structured educational content. Rule-based systems (4/17, 23.5%) appear primarily in counseling contexts where greater conversational flexibility is needed while maintaining therapeutic boundaries. This distribution suggests that different therapeutic contexts require different levels of conversational sophistication and autonomy. The continued use of Wizard of Oz techniques (3/17, 17.6%) in recent studies indicates that fully autonomous digital humans may not yet be ready for complex therapeutic interactions, particularly those requiring real-time clinical judgment. This approach allows researchers to explore the therapeutic potential of digital humans while ensuring participant safety and intervention quality. The innovative use of recorded playback in CFT [35] demonstrates how back-end intelligence can be tailored to specific therapeutic mechanisms, using participants’ own recorded behaviors to create personalized therapeutic content.

The design of digital humans for depression support requires balancing multiple factors: Visual appearance needs to be culturally adapted to the target population and matched to platform capabilities, and within these constraints, should be realistic enough for credibility while avoiding uncanny valley effects. Interaction design should prioritize natural speech input where feasible but adapt to environmental constraints. Back-end intelligence approaches must be selected based on therapeutic context, with scripted systems providing safety and standardization for assessments, while rule-based and Wizard of Oz techniques enable more flexible therapeutic interactions. Ultimately, successful digital human implementation depends on aligning design choices with specific therapeutic goals, technological capabilities, and user needs rather than pursuing technological sophistication for its own sake.

The therapeutic alliance, characterized by a collaborative and trusting relationship between a health provider and a client, is a cornerstone of effective psychotherapy [61]. Traditionally, this alliance is built through face-to-face interactions, where empathy, understanding, and mutual respect foster a sense of safety and connection. With the advent of digital humans in therapeutic settings, there is a growing interest in understanding how these virtual entities can contribute to building and maintaining a therapeutic alliance [62].

In this scoping review, multiple studies [23,36,37,39] have demonstrated the potential of digital humans to establish trustworthy and comfortable relationships with clients. These findings underscore the viability of digital humans as supportive agents in mental health care. Several key factors contribute to the successful establishment of therapeutic alliances with digital humans. First is empathy and emotional responsiveness. The ability to empathize is crucial in building a therapeutic alliance [63]. Digital humans, as embodied virtual agents, can combine facial expressions, body movements, and emotional narration to convey emotions accurately. In the study by Ring et al [37], participants felt that their feelings were understood because the digital human provided appropriate feedback. Similar results were observed in the experiment by Philip et al [39], where participants also reported a sense of being understood. These findings highlight the importance of empathetic interactions in fostering a strong therapeutic alliance. Another critical factor is the nonjudgmental presence of digital humans. This presence is created through the careful design and control of the content expressed by the digital human. The interactive content of digital humans applied in all of the 17 included studies was designed or evaluated by experts before being applied in experiments. Although this approach limits the range of user interactions, it ensures the safety and appropriateness of the interactions. By providing a nonjudgmental and safe space, digital humans can help clients feel more comfortable and open, thereby strengthening the therapeutic alliance. In addition, compared to face-to-face interactions, digital humans offer customizable support tailored to the individual’s personality and needs. This personalized attention can make people with depression feel valued and understood, increasing their motivation to engage in therapy and fostering a deeper sense of trust.

Digital humans have demonstrated significant potential in building therapeutic alliances with individuals experiencing depression. By providing empathetic, consistent, and personalized support, digital humans can help overcome barriers related to stigma, availability, and accessibility.

While this scoping review demonstrates the promising potential of digital humans in depression management, implementation raises critical ethical concerns requiring careful attention. Privacy and data security present immediate risks, as digital human systems collect highly sensitive information, including verbal responses [35], behavioral patterns [39], and potentially video or audio recordings during vulnerable therapeutic moments [37]. Many reviewed studies provided limited detail about data management practices, yet these systems often rely on cloud-based infrastructure and third-party services that create multiple exposure points for sensitive mental health data. Biometric information, such as voice recordings and facial expressions, is inherently difficult to anonymize and could potentially reidentify individuals even after traditional identifiers are removed. Future implementations should ensure robust encryption, secure storage protocols compliant with relevant regulations, transparent disclosure of data practices, and clear policies regarding third-party access and data retention.

The accessibility advantages of digital humans paradoxically create risks of overreliance and inappropriate use. Recent news has shown that individuals may incorrectly view these tools as complete replacements for human therapists rather than complementary supports [64], potentially delaying access to necessary human intervention for severe depression, suicidal ideation, or complex presentations. Current autonomous systems have limited capacity to recognize and appropriately respond to acute crises, and users may form parasocial relationships with digital humans without recognizing fundamental limitations compared with human therapeutic relationships. Clear communication about system boundaries is essential, alongside robust escalation mechanisms to human providers when needed. Additionally, questions of accountability remain unresolved—when adverse outcomes occur, responsibility distributed among developers, health care providers, and users requires clear frameworks that protect vulnerable populations while enabling innovation. Equity and access concerns threaten to undermine the democratizing potential of digital humans. Implementation requires technological access (appropriate devices, reliable internet, and digital literacy) that populations at risk for depression—including low-income individuals, older adult persons, and those in rural areas—may lack. Only 25% (5/20) [23,25,33,34,38] of reviewed studies occurred in everyday environments, suggesting limited real-world accessibility evidence. Without deliberate policy intervention ensuring equitable access through public health systems and subsidized programs, these technologies risk becoming available primarily through private markets, widening rather than closing mental health care gaps.

Moving forward, ethical implementation requires privacy-by-design approaches with comprehensive data protection, transparent communication of system limitations and appropriate use cases, inclusive development ensuring diverse representation in training data and validation studies, and hybrid care models positioning digital humans as adjuncts rather than replacements for human providers. Regulatory frameworks must establish appropriate validation standards and postmarket surveillance requirements, while sustainable funding models must prioritize access for underserved populations. As this technology advances beyond proof-of-concept demonstrations, the field should evaluate not only what digital humans can do, but what they should do, ensuring that the pursuit of innovation maintains focus on user well-being, autonomy, safety, and equitable access across all populations who might benefit from mental health support.

The review highlights that effective digital human systems are grounded in collaborative design processes between clinicians and developers. To ensure therapeutic relevance and usability:

This interdisciplinary co-design ensures that systems are both clinically sound and technically viable.

The deployment of digital humans in mental health care raises important ethical considerations that require careful attention during study. Based on the review findings, we suggest:

Successful implementation of digital humans in depression care requires strategic planning across multiple dimensions: