This National Institute of Mental Health–funded (R41MH131229) study aimed to recruit a final sample of 40 clinicians. Of the 42 participants enrolled in the therapist training trial, one withdrew after consenting due to scheduling difficulties. The remaining clinicians (N=41) completed all stages of the training trial, including follow-up. Clinicians were recruited from a variety of settings between November 2023 and April 2024. Recruitment methods included convenience sampling (recruiting from the pool of clinicians at Bradley Hospital), offering training to clinicians on a waitlist to participate in nonexperimental CBT training at Bradley led by the first author, and targeted email advertising to social workers, private-practice clinicians, and providers affiliated with a local mental health practice. The eligibility criteria for the study included a bachelor’s level of education (or higher), the ability to come to Bradley Hospital for in-person experiential training, and the ability to deliver clinical care to patients in a professional capacity.

The majority of the therapist participants (37/41, 90%) were White and non-Hispanic. Four out of 41 (10%) therapists reported Hispanic or non-White identities. A majority (35/41, 85%) also identified as female. Participants had a variety of degree types, including doctorate (n=5), master’s (n=27), and bachelor’s-level (n=9) degrees. Licensure types include clinical psychologists (n=4), mental health counselors (n=13), social workers (n=13), and either graduate students or other (non-licensed) clinical roles (n=11).

Study procedures and informed consent documentation were approved by the Brown University Health Institutional Review Board (IRB00004624) in July 2022. All participants provided written consent after meeting with study staff to review the consent documents. As compensation for time spent engaged in study activities, participants received continuing education credits and payment in the form of either cash or an Amazon gift card. Both continuing education credits and monetary compensation (US $300 total) were awarded to participants after they completed all parts of the study. To protect the privacy and confidentiality of participant information, identifying information from the consent process was stored on a secure password-protected server, and all analyses were conducted using a fully deidentified dataset. This is not a secondary analysis, and consent was provided for all reported data. There is no identifying information presented in the analyses or supporting tables and figures. No adverse events or side effects were reported at any point during or after the trial. Study details are reported using CONSORT (Consolidated Standards of Reporting Trials) 2025 (Checklist 1) and CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) guidelines [40] (Checklist 2).

Statistical analyses were carried out using SAS version 9.4 1M7. Knowledge, self-efficacy, and TBES were assessed longitudinally and analyzed via binomial generalized estimating equations (GEE). SUS, VP Eval, and TARS were assessed only at the conclusion of the study and analyzed using a binomial generalized linear model. Both sets of analyses used the binomial distribution to constrain the model space to the range of measures being used. The scores were first shifted to have a minimum score of zero by subtracting the minimum score possible. This was used as the binomial number of “successes,” and the range of the score was used as the number of “trials.” Classical sandwich estimation was then used to better incorporate any differences in empirical variances and covariances. Both raw and adjusted (where appropriate) P values were presented, as well as unadjusted 95% CIs. For GEEs, 16 different hypotheses were tested using orthogonal linear estimates: (1) changes from baseline conditions in each condition (8 total), (2) high- and low-immersion training conditions compared at each timepoint cross-sectionally (4), and (3) condition comparisons in changes from baseline for each follow-up (4). The Holm test was used to calculate adjusted P values, maintaining familywise alpha at .05 across all 16 hypothesis tests in GEEs. For all other models, only 1 hypothesis was tested comparing the conditions. There was no attrition after randomization and participants were required to provide values for all survey items except open-text responses; imputation techniques were not used given the lack of missing data.

The high- and low-immersion training conditions did not exhibit statistically significant differences on measures of exposure knowledge (t₁₁₇=1.75, P=.08), exposure self-efficacy (t₁₁₇=1.90, P=.06), and attitudes about exposure therapy (t₁₁₇=0.72, P=.47) before initiating the workshop training, indicating both groups had similar ratings on key variables prior to initiating the training intervention (Figure 3) [40].

There were no statistically significant differences between the groups on measures of exposure knowledge (t₁₁₇=0.93, P=0.35), exposure self-efficacy (t₁₁₇=1.40, P=.17), and attitudes about exposure (t₁₁₇=0.00, P=.998) following the completion of the didactic portion of training, which suggests both groups demonstrated similar amounts of change in these key variables from baseline to postdidactic timepoints.

The headset and computer groups exhibited significant improvement in exposure knowledge (t₁₁₇=3.16, P=.002 and t₁₁₇=2.81, P=.006, respectively), exposure self-efficacy (t₁₁₇=7.37, P<.001 and t₁₁₇=4.75, P<.001, respectively), and attitudes about exposure (t₁₁₇=5.68, P<.001 and t₁₁₇=3.32, P=.001, respectively) from baseline to postdidactic timepoints.

There were no significant differences in exposure knowledge (t₁₁₇=1.41, P=.16), exposure self-efficacy (t₁₁₇=1.38, P=.17), and attitudes about exposure (t₁₁₇=0.11, P=.92) at the end of the workshop training.

The headset and computer groups demonstrated significant improvement in exposure self-efficacy (t₁₁₇=5.03, P<.001 and t₁₁₇=3.58, P<.001, respectively) and attitudes about exposure therapy (t₁₁₇=3.56, P<.001 and t₁₁₇=2.5, P=.01, respectively) from post-didactic to the end of the workshop timepoints. Neither group demonstrated significant change in exposure knowledge from post-didactic to the end of workshop (t₁₁₇=0.19, P=.853 and t₁₁₇=0.33, P=.74, respectively). Taken together, these findings indicate that the experiential portion of training significantly improved therapists’ perceptions of self-efficacy in delivering exposure therapy and their attitudes toward exposure therapy, but it did not significantly improve their knowledge of exposure therapy beyond ratings taken after the completion of didactic activities. Mean knowledge scores at T1 and T2 (Table 2) remained well below the maximum possible score of 12, indicating that the absence of additional knowledge gains from post-didactic to post-experiential was not attributable to a ceiling effect. Instead, this pattern is consistent with the design of the training, in which no new didactic exposure content was introduced during the VR phase (see Figures 4-6).

There were no significant differences in exposure knowledge (t₁₁₇=1.44; P=.15), exposure self-efficacy (t₁₁₇=1.38; P=.17), and attitudes about exposure (t₁₁₇=0.39; P=.70) after a 1-month follow-up period, indicating comparable maintenance of learning between conditions.

The headset and computer groups exhibited durable maintenance of gains during the 1-month follow-up period on measures of exposure knowledge (t₁₁₇=0.17; P=.87 and t₁₁₇=0.49; P=.63, respectively), self-efficacy (t₁₁₇=1.26; P=.21 and t₁₁₇=1.50; P=.14, respectively), and attitudes about exposure (t₁₁₇=0.89; P=.38 and t₁₁₇=0.94; P=.35, respectively).

Ratings of system usability using the SUS measure were statistically significantly higher for the computer group relative to the headset group (t₃₉=−2.87, P=.007); however, usability for both groups (HMD: mean 86.2, SD 9.1, 95% CI 81.68-89.75; desktop: mean 93.4, SD 6.7, 95% CI 89.87-95.80) was rated highly, with both modalities ranking in the upper 96th percentile of usability according to established norms [41]. Therapist perceptions of the virtual patient’s authenticity, degree of engagement, and overall utility as a learning experience, as assessed by the VP Eval measure, were similarly positive in both groups (t₃₉=0.08; P=.93).

Ratings of the training processes and their acceptability using the TARS measure were positive. The groups exhibited nearly identical scores on ratings of the training’s acceptability, which assess general acceptability, perceived effectiveness, and appropriateness of the training; however, the computer group reported significantly more positive ratings of training processes (t₃₉=−2.02; P=.05) relative to the headset group, which assessed the extent to which the teaching approach felt helpful.

This study tested the feasibility and preliminary effectiveness of a novel exposure therapy training program applying VR technology within an E2E framework. Two delivery modalities with varying degrees of immersion were piloted as different experiential doses of E2E practice to evaluate each method’s usability and impact on key training targets. Clinicians were randomized to training programs of low immersion (desktop computer) and high immersion (HMD) and assessed for changes in exposure knowledge, self-efficacy, and attitudes about exposure therapy.

Our primary outcomes demonstrate that virtual technology is highly usable and effective in engaging relevant therapist-level learning variables during exposure therapy training. Participants rated both virtual programs highly on the SUS, which measures the programs’ ease of use, intuitiveness, and functionality. Likewise, the programs received high scores on the VP evaluation scale, indicating that the virtual patient felt realistic to the therapists. This is especially promising as one challenge in conventional training is the limited ability of role-play to capture the real experience of delivering exposures [12].

Two of the primary training metrics in this study were therapists’ perceptions of their efficacy in delivering exposure therapy and their attitudes about the safety and tolerability of exposure for their clients. Both metrics have been previously identified as therapist-level barriers that limit the frequency and quality of exposure delivery in typical practice [18]. Participants showed significant improvements in the self-efficacy measure, which has been shown to have positive predictive validity for future exposure use [20]. The improvements in TBES scores between T1 (after didactic training) and T2 (after SET-VR) demonstrate that virtual experiential learning further reduced negative beliefs and increased comfort with performing exposures. Notably, these findings were sustained at the 4-week follow-up, indicating potential long-term gains from virtual training. Participants did not show changes in levels of knowledge between T1 and T2. This is to be expected as the SET-VR program focuses on experiential learning with the assumption that the therapist has the necessary theoretical underpinnings. However, our results support virtual experiential learning as a promising option for further development to address key barriers in the current models of exposure therapy training. In contrast, exposure knowledge was included primarily as a control-like outcome to verify that both groups benefited similarly from the didactic workshop and entered the VR phase with comparable background knowledge. This approach is consistent with prior work showing that knowledge about exposure is necessary but far from sufficient to change exposure use or quality, whereas self-efficacy and negative beliefs represent more central barriers to implementation [9].

Interpreted through the E2E lens, these findings suggest that VR-based experiential practice can function as an exposure intervention directed at therapists’ own anxious beliefs about exposure. The additional reductions in negative beliefs and gains in self-efficacy from T1 (post-didactic) to T2 (post-experiential) indicate that repeatedly “doing” exposure with a distressed virtual patient can provide corrective learning beyond what is achievable through didactic instruction alone, even when knowledge is already adequate. This psychotherapy-specific mechanism focus—targeting therapist beliefs that uniquely impede exposure implementation—was central to the rationale for using VR in this trial.

The results from the direct comparison of the 2 delivery conditions addressed our central question about how much immersion is needed for E2E training and did not support our initial hypothesis that the high-immersion condition (HMD) would outperform the low-immersion condition (desktop computer). Participants in both conditions reported comparably high values for the replication of real-life scenarios on VP Eval and satisfaction with the training on the TARS. While both groups gave high ratings on usability (SUS), the low immersion group’s rating was higher at a statistically significant level. As for training outcomes, both conditions improved self-efficacy and negative attitudes without statistically significant between-group differences. Taken together, this finding shows that, at least within this feasibility study, the use of HMD to deliver the virtual training was not superior to the use of a desktop computer in addressing key therapist-level targets, and the lower immersion method may have higher usability. From an E2E perspective, this pattern suggests that, in the context of a brief single-session training, a lower-immersion desktop may already provide sufficient presence and emotional engagement with a distressed virtual patient to facilitate learning about the safety and tolerability of exposure. Higher-immersion HMD may yield incremental benefits only under conditions of greater training “dose” (eg, more repetitions, more complex scenarios) or in subgroups of learners; this is a question that will require a larger, mechanism-focused trial.

In the context of the larger medical education literature, there may also be a task-technology mismatch between virtual exposure therapy delivery and high-immersion VR headset. Prior studies demonstrating a positive correlation between immersion factors and medical education learning outcomes have shown that the effect is most significant in spatial and procedural tasks, such as surgical procedures [42] and physical examinations [43]. Conducting exposure therapy does have procedural elements, like the sequential progression of actions based on branching decision points. However, our participants’ tasks are different from physical procedures in that they do not need a high degree of spatial awareness or physical interaction with the virtual environment, which would be the main advantages of an HMD [44]. Thus, it is possible that visual and auditory stimuli through a desktop computer provided the same necessary input for experiential learning of delivering exposures.

The incremental benefits to learning provided by higher immersion could have been negated by the increased cognitive demands that arise from operating an HMD. Task-irrelevant sensory information or the mental effort required to navigate the less familiar HMD interface could have been distractors from the key experiential task. This would be consistent with the cognitive load theory framework, which states that extraneous cognitive demands reduce mental resources needed for optimal learning [45,46]. Additionally, methodological factors, such as participation selection bias and sample size, may have confounded the detection of between-group differences. One of the most well-studied benefits of high immersion VR is increased user engagement with the material [44]. Given that participants are a self-selected group who are motivated in exposure therapy training, this effect may have been overshadowed. It was anecdotally observed by the study team that participants in the HMD group demonstrated more consistent engagement with the exercise. It is possible that over a longer duration of training, differences in engagement could impact learner outcomes.

Overall, our findings present a basis for further investigation into how immersion functions as an experiential dose parameter within E2E-based psychotherapy training, as well as an opportunity for flexible implementation of VR technology based on factors such as technology literacy, cost-effectiveness, and learner preferences. Most existing studies of VR training in mental health focus on skills, knowledge, and attitudes for specific diagnoses (eg, depression, psychosis) [29]. In contrast, while this study involves a virtual patient with anxiety disorders, its primary aim is to enhance learner outcomes with a specific behavioral health technique (exposure therapy). In this regard, very little research has been done to evaluate procedural learning in VR for behavioral therapy. One notable review identified technology-based methods for training behavioral counseling skills [27]. It found that a mix of virtually created patients and filmed actors was being used. A majority of available outcomes were related to using simulations as opportunities for the repetition of single skills (affirmation, reflective listening, change talk, psychoeducation) [27,47,48], rather than aiming to deliver experiential learning for high-stress clinical scenarios or to address provider beliefs about a treatment modality.

To our knowledge, this is the first study to leverage virtual technology to augment exposure therapy training. The current norm of didactics-focused training is effective in instilling knowledge and, to some extent, improving clinician self-efficacy and negative beliefs [49], as demonstrated even in our single-session training intervention. However, the literature also shows that provider-level barriers continue to limit the optimal dissemination of exposure therapy [13], despite it being one of the most effective tools in psychiatry [18]. Thus, our finding that the supplementation of didactic learning with virtual experiential learning further improves self-efficacy and negative beliefs about exposure therapy suggests a promising area for further investigation to optimize the standardization and scalability of exposure training. This is particularly notable as VR technology in itself offers logistical advantages (accessibility, scalability, reproducibility), the ability to replicate rare or high-risk situations in a safe training space, and near-limitless customizability [50].

Our study design contains notable limitations. There were clear imbalances in the participant demographics. A large majority (90%) of participants were White, non-Hispanic, and female (85%). The gender imbalance may be particularly pertinent, as there have been heterogeneous results regarding the relationship between gender and VR experiences such as immersion, emotional response, and task performance [51]. There is also the potential for selection bias in the recruited participants. While technology-based training interventions are generally well accepted [27], there may be clinicians who prefer other modalities. Those who opted to join this study may be more interested in innovation or technology, which could have influenced their experience or learning outcomes. Additionally, participants were self-selected based on their motivation for exposure therapy training. Although it was not observed in our study, it is possible that cyber sickness and the initial learning curve of new digital platforms may present barriers to engagement in a more general clinician population.

This feasibility trial is designed to assess the effectiveness of SET-VR in addition to didactic training; thus, it is not possible to draw direct comparisons between SET-VR and standard didactic training alone. While a variety of carefully selected metrics were used to assess training efficacy, there are no direct measures of how the quality or frequency of exposures delivered by clinicians changed after SET-VR. Our most reliable proxy is the improvements in exposure self-efficacy, which has previously been shown to have predictive validity in the future likelihood of using exposure therapy [20]. Finally, the current version of SET-VR does not use novel algorithms or adaptive rendering techniques; instead, it uses a controlled, pre-programmed decision structure to standardize the training experience. While this limits technological novelty, it was appropriate for a feasibility trial focused on mechanism-relevant learning outcomes.

This study offers directions for further investigations. In the current trial, the SET-VR architecture intentionally relied on a relatively simple, preprogrammed decision tree and automated SUDS trajectory, rather than more complex adaptive algorithms, to maximize standardization and internal validity for testing immersion and belief change. Future iterations are planned to incorporate more complex interaction logic, including natural language processing–based virtual patient responses and richer affective behaviors. Future iterations will also include a longer training course to see if empirical observations of differences in engagement between immersion levels lead to significant differences in learning. As with other training research, it will be important to evaluate how performance in the exposure simulation sessions correlates with behaviors in actual clinical practice. Noteworthy outcomes could include an increased likelihood of using exposures overall, as well as higher fidelity to the principles of effective exposure therapy [18]. To that end, a larger training trial with practicing clinicians is planned, which will compare decisions during the simulation against coded video tapes of exposure delivery with patients after training to evaluate how performance during training aligns with future delivery.

The adaptation of VR technology into clinical training offers new opportunities to overcome longstanding limitations in the accessibility, standardization, and experiential quality of exposure training. This study is innovative in applying an explicit “E2E” framework to VR, using simulated exposure sessions to directly target therapists’ anxious beliefs about exposure, and in its examination of immersion as a dose parameter rather than assuming that “more immersive” is inherently better. In contrast to much of the existing VR training literature, which has focused primarily on knowledge and procedural skill acquisition, our findings highlight that both desktop and HMD implementations can produce meaningful changes in self-efficacy and negative beliefs, which are mechanisms that are uniquely tied to exposure underuse in routine care. The lack of clear incremental benefit for HMD over desktop VR suggests that lower-immersion, lower-cost formats may be sufficient to deliver an effective experiential dose of exposure practice, which has important implications for scalability and implementation in community and academic settings. Taken together, these results provide preliminary but encouraging evidence that E2E-guided VR training can enhance conventional didactic training and may be a promising next step in optimizing the standardization and effectiveness of exposure training.