The study began with a preparation meeting (about 80 Min) including an information brochure and the possibility to ask questions. The investigator (A.R.) explained eligibility criteria, emphasizing voluntary participation and the option to withdraw without giving reasons. After signing consent forms, participants downloaded the 7Mind^® app (7Mind^®, 2024) and received instructions. Those in group meetings received organizational details, while VR participants were introduced to the Meta Quest 2 headset and Horizon Workrooms software (Meta, 2024). Loaned headsets allowed familiarization with the VR software a week before the course. Pretest data were collected, participants were randomly assigned to conditions, completed the MBI (groups: active control, AC; video call, VC; virtual reality; VR) or waited (passive control group, PC), and then took post-test measures before compensation. At the end of the investigation period, a 45-min follow-up meeting was held.

The MBI course by 7Mind^® was based on the principles of mindfulness-based programs (Crane et al., 2017). It included eight 60-min audio modules on stress, mindfulness exercises, and meditation. After each session, participants received a summary handout and were instructed to practice independently. The course required weekly module completion and passing a quiz to advance (7Mind^®, 2024). The MBI was accompanied by weekly online meetings in small groups of five to seven participants (i.e., 8 × 60 min) within two intervention conditions: either as 2D video call meetings (using Webex, VC group) or in a virtual 3D meeting room using VR glasses (VR group, see below). The group meetings served to structure and elaborate the 7Mind^® podcasts and to facilitate experience-based group exchange. The course instructor (A.R.) was psychologist by training (M.Sc.) and had teaching experiences with university students.

Participants were randomly assigned to one of four conditions. However, minor irregularities in randomization could not be avoided for organizational reasons (see SM1 in Supplementary material for details). The passive control group (PC) only completed pre- and post-test assessments (n = 32). The active control group (AC) completed the MBI (n = 35), but without weekly group meetings. Two additional groups supplemented the MBI with weekly online meetings as described above. Specifically, the video call group met via a 2D video platform (VC, n = 31), while the VR group used Meta Quest 2 VR glasses to meet in a 3D environment on Horizon Workrooms (VR, n = 29), featuring a virtual landscape with avatars around a table (for demonstration see SF1 in Supplementary material). Apart from this, the group meetings were held in the same way in both intervention groups. A flowchart of the study design is shown in Figure 1.

Participants were recruited via the Online Recruitment System for Economic Experiments (ORSEE) (Greiner, 2015), and gave written informed consent in accordance with the Declaration of Helsinki before taking part. The study was approved by the local ethics committee (University of Lübeck). All participants confirmed that they were students at the University of Lübeck, of legal age and in good health (i.e., no former or current mental or physical illnesses such as anxiety disorders or cardiovascular problems), consumed alcohol and nicotine only in moderation, had a good sense of balance and were physically healthy, and that they were currently not taking any prescription medication (except birth control pill). A total of 130 participants were recruited via email. Two participants withdrew from the study due to time constraints. One participant was excluded due to missings across all measurements. The final sample with minimal missing data and three dropouts at second measurement point (see Supplementary material) contained a total of 127 students participated (105 female, 21 male, 1 unspecified), with a mean age of 24.06 years (SD = 5.02, range = 18–52). Of these, 72 were psychology students, and 33 had a Bachelor's degree, averaging 4.21 semesters of study (SD = 2.63).

Nineteen subjects demonstrated elevated levels of mental stress, identified by the Brief Symptom Checklist (BSCL, see below) pretest scores (T ≥ 63 indicating high-risk subjects; Franke et al., 2015). However, we did not remove these participants from the final sample, assuming these were acceptable variations within the normal range. In the intervention groups (n = 88), 26 reported prior VR experience, though none owned a headset. Participants chose either €100 or €75 plus five hours of course credits as compensation. Additionally, all subjects had 12 months of free access to the 7Mind^® app (7Mind^®, 2024); the PC group received access only after the end of their study participation.

We report all collected variables, including those not analyzed, in accordance with JARS guidelines (Appelbaum et al., 2018). While not part of the main analyses, additional self-generated variables were assessed to evaluate participants' experiences in the MBI groups (AC/VC/VR) and included open-ended responses (e.g., perceived usefulness or potential barriers of the VR glasses). Free-text responses are summarized in the Supplementary material. The Supplementary material and OSF material provides additional methodological details, sample characteristics, and supplementary analyses. These materials enhance transparency and allow for replication.

All measures contained self-reported Likert scales. Unless otherwise specified, mean values were calculated for all scales. Cronbach's alpha was used to assess internal consistency; for pre- and post-test measurements, separate alpha values were reported.

With respect to BSCL, several outliers were identified based on QQ normality plots. Therefore, we used an alternative operationalization based on a “tendency toward stress” approach (counting items > 0; see Franke, 2017). This adjustment reduced outliers in QQ plots, mitigating severe violations of parametric assumptions and allowing us to avoid removing data points. Missing data were minimal and handled by mean imputation at the item level within each scale, consistent with recommendations for handling small proportions of MCAR data (Tabachnick and Fidell, 2019).

For hypothesis testing, we used Linear Mixed Models (LMM) and Ordinary Least Squares (OLS) regression models, reporting unstandardized regression coefficients (b), standard errors (SE), p-values, and effect sizes (R² as well as Cohen's d for mean differences). Given only minor deviations from normality, we relied on parametric significance estimates for hypothesis testing. To assess robustness, we also computed bias-corrected and accelerated (BCa) bootstrap confidence intervals (95% CI, 5,000 samples) for (fixed) effects (LMM/OLS). Furthermore, we calculated parametric bootstrap confidence intervals (95% CI, 5,000 samples) for random effects (LMM). We report only the BCa confidence intervals as an additional robustness check, even in cases where these intervals may include zero. Data were analyzed using R, version 4.2.2 (R Core Team, 2022), with lme4 (Bates et al., 2015) for mixed-effects modeling, ggplot2 (Wickham et al., 2016) for visualization, mediation (Tingley et al., 2014) for conducting mediational analyses, pwr (Champely et al., 2020) and simr (Green and MacLeod, 2016) packages for a post-hoc power and sensitivity analyses.

A post-hoc sensitivity analysis assessed the model's power to detect the strongest observed interaction effect (PC vs. AC/VC/VR × time on self-compassion: self-kindness) using pwr package. For the obtained effect size of b = 0.10 [f² = 0.071; representing a small-to-moderate effect size according to Cohen's (1988) guidelines] with df = 126 and α = 0.05, the analysis yielded a power of 1 – β = 0.85, indicating sufficient sensitivity for detecting the hypothesized interaction (see Subscale's Analyses). In addition, to evaluate the sensitivity of the model testing, a simulation-based power analysis was conducted using simr package. Based on the same model (PC vs. AC/VC/VR × time on self-compassion: self-kindness) with 251 observations, the minimum detectable effect size (MDES) was estimated. With 2,000 simulations at α = 0.05, the model achieved 81.05% power (95% CI: [79.26%; 82.75%]) to detect an effect estimate of the interaction term PC vs. AC/VC/VR × time of b = 0.10. Thus, the analysis confirms that the empirical effect was detected with sufficient statistical power.

To test if baseline characteristics (e.g., sociodemographic and pretest scores) were independent of intervention conditions, χ² tests, ANalyses Of VAriance (ANOVAs), and Ordinary Least Squares (OLS) regressions were conducted. χ² tests showed independence for gender [χ²₍₃₎ = 1.603, p = 0.659], study program [χ²₍₃₎ = 0.885, p = 0.829], study degree [χ²(3) = 3.481, p = 0.323], and VR experience [χ²₍₂₎ = 4.497, p = 0.106]. ANOVAs showed no significant differences by condition for age [F_{(3, 123)} = 1.261, p = 0.291] and semester [F_{(3, 121)} = 0.157, p = 0.925]. OLS regressions for pretest scores indicated overall minor dependencies for BSCL subscale depression [F_{(3, 121)} = 3.136, p = 0.028], but no systematic differences (all ps ≥ 0.05) for remaining outcomes. Minor trends emerged for loneliness and life satisfaction (0.05 ≥ps < 0.10). High-risk subjects (derived from BSCL: T ≥ 63; Franke et al., 2015), were evenly distributed across conditions [χ²₍₃₎ = 2.499, p = 0.518]. Thus, baseline characteristics were considered sufficiently independent of intervention conditions. Additionally, the dependent variables showed the expected pattern of intercorrelations. For details refer to ST1 (Supplementary material).

To test our hypotheses, hierarchical Linear Mixed Models (LMM) were conducted with subjects as random intercepts. Model 1 included intercepts only, while Model 2 added time (pretest vs. post-test, coded 0/1), and Model 3 included three orthogonal Helmert contrasts for intervention conditions: (1) PC vs. AC/VC/VR (PC: −3, others: 1); (2) AC vs. VC/VR (AC: −2, others: 1); and (3) VC vs. VR (VC: −1, VR: 1). These unscaled contrasts reflect the stepwise structure of the intervention manipulation. Coefficients are based on raw contrast weights; hypothesis tests remain unaffected by contrast scaling (Cohen et al., 2013). The last saturated Model 4 included the three interaction terms [time × contrast (1), time × contrast (2), and time × contrast (3)] for testing the hypotheses H.1, H.2, and H.3. For model comparison, χ²-Tests were conducted for testing incremental increases in variance explanation.

Analyses began with overall scales and proceeded to subscales. The intercept-only model (Model 1) showed substantial within-person variability across measures (ICC range: 0.30–0.82). Main effects by time in Model 2 (pre- vs. post-test) showed reductions in perceived stress [PSS-10: b = −0.17, SE = 0.04, 95% Boot CI = (−0.274; −0.058), Boot SE = 0.06, d = −0.33, p < 0.001] and increases in self-compassion [b = 0.17, SE = 0.04, 95% Boot CI = (0.069; 0.275), Boot SE = 0.05, d = 0.37, p < 0.001], self-efficacy [b = 0.09, SE = 0.03, 95% Boot CI = (0.024; 0.166), Boot SE = 0.04, d = 0.28, p = 0.002], and life satisfaction [b = 0.25, SE = 0.06, 95% Boot CI = (0.107; 0.395), Boot SE = 0.07, d = 0.39, p < 0.001]. Subscale time effects yielded reductions in stress [e.g., helplessness: b = −0.17, SE = 0.05, 95% Boot CI = (−0.301; −0.046), Boot SE = 0.06, d = −0.29, p = 0.001] and loneliness [social isolation: b = −0.10, SE = 0.04, 95% Boot CI = (−0.187; −0.003), Boot SE = 0.05, d = −0.23, p = 0.011] and increases in self-compassion subscales [e.g., self-kindness: b = 0.16, SE = 0.06, 95% Boot CI = (0.016; 0.307), Boot SE = 0.08, d = 0.23, p = 0.011; isolation: b = 0.18, SE = 0.07, 95% Boot CI = (0.014; 0.340), Boot SE = 0.08, d = 0.24, p = 0.007]. Trends were observed for reduction in perceived stress [PSQ-20: b = −0.02, SE = 0.01, 95% Boot CI = (−0.053; 0.011), Boot SE = 0.01, d = −0.15, p = 0.096; demands: b = −0.03, SE = 0.02, 95% Boot CI = (−0.072; 0.010), Boot SE = 0.02, d = −0.17, p = 0.057]. The following models (Model 3) were build up as incremental references for the last models (Model 4) testing the hypotheses at once.

The initial hypothesis (H.1) was based on the premise that our MBI would be effective across all intervention groups (AC/VC/VR) when compared to a passive control group (PC). This was confirmed with significant effects on reducing stress and loneliness, yet increasing self-compassion. Our findings align with those of previous studies, which have demonstrated that MBIs can enhance mental and physical wellbeing by mitigating stress (McConville et al., 2017; Grossman et al., 2004). Similarly, the reduction of loneliness can be considered a beneficial outcome in line with previous studies (Besse et al., 2022). Our MBI was found to enhance the tendency to treat oneself with kindness and compassion, which should, in turn, contribute to an improvement in the wellbeing of those who participated in the course. Moreover, the initial findings were corroborated, indicating that app-based mindfulness training can also have beneficial effects.

Subsequent contributions have also demonstrated the stress-reducing impact of an app-delivered mindfulness intervention. While Karing (2024), Huberty et al. (2019), and Riley et al. (2022) demonstrated positive outcomes of app-delivered mindfulness interventions, their designs lacked features such as structured group meetings or immersive technologies. For instance, Schwartz et al. (2023) reviewed randomized controlled trials of mindfulness-based mobile apps and found that these apps improved wellbeing, but the absence of group dynamics and contextual support often limited their impact. Hwang et al. (2022) evaluated a smartphone-based stress reduction app and reported positive effects on stress levels, although the study did not incorporate group support or immersive features that might further enhance user engagement. In contrast, our study integrated weekly online group meetings and a VR environment, aligning more closely with the immersive frameworks discussed by Arpaia et al. (2021). Their narrative review highlights that VR-enhanced mindfulness-based applications may amplify emotional and sensory engagement, though few empirical studies have explored this in student populations. Furthermore, Xu et al. (2022) conducted a mixed-methods review and found that virtual mindfulness interventions positively impacted wellbeing, but most studies lacked rigorous mediation analysis or comparison of delivery formats (e.g., video vs. VR). Our study adds value by addressing these gaps—evaluating mindfulness as a mediator and assessing VR immersion subcomponents.

The second hypothesis (H.2) posits that regular online group meetings (VC/VR) will enhance the efficacy of the MBI relative to the active control group (AC). These meetings may also facilitate stress reduction through social interactions and collective self-efficacy (Cruwys et al., 2020; Yusufov et al., 2019). However, no significant additive effects were found for the group meeting conditions (VC/VR) compared to the active control condition (AC). Unexpectedly, the AC showed greater stress reduction than group meetings, yet consistent with some studies showing limited benefits of video-based mindfulness interventions over active control groups (Moulton-Perkins et al., 2022).

In the associated exploratory research question (E.1), we postulated that mindfulness acted as a mediator, given that this represents a central mechanism of the mindfulness-based training. All intervention groups (AC/VC/VR) showed significantly higher mindfulness at post-test (cf. Karing, 2024; Huberty et al., 2019), with group meetings (VC/VR) enhancing mindfulness more than podcast-only (AC), which is in line with prior studies (Visted et al., 2015). This mindfulness increase, often linked to stress reduction (Keng et al., 2012), may reflect a suppressor effect, where group meetings elevated social stress in some subgroups, dampening the actually positive outcomes of the MBI. Indeed, such a pattern was observed in studies of group-based stress reduction trainings, where individual levels of group identification and social support varied (see Imel et al., 2008; McKimmie et al., 2020). Future studies should investigate whether and under what conditions certain subgroups could benefit from (online) group meetings.

In our last hypothesis (H.3), we proposed that group meetings enhanced with a VR environment would provide additional benefits over those with conventional VC software. Indeed, VR environments could enhance the effects of a mindfulness-based training over traditional approaches (Ma et al., 2023) possibly by intensifying sensory experiences (Turner and Casey, 2014) and strengthening motivation and adherence (Jiang and Fryer, 2024; Modrego-Alarcón et al., 2021). However, we did not observe any direct effects in our study. Free-text comments (SM3 in Supplementary material) noted some VR-related burdens (e.g., discomfort, headache, technical issues), possibly explaining the lack of additional VR effects, similar to challenges noted in earlier work on VR discomfort (Jensen and Konradsen, 2018). However, these rather subjective comments should be interpreted with caution.

The second exploratory research question (E.2) examined the potential association between the intensity of immersion in the VR environment and our outcome variables. Although the global measure of VR immersion demonstrated no correlation with the outcomes, an examination of single sub-dimensions of VR immersion revealed contrasting results. Higher physical presence was linked to reduced stress, whereas higher social presence had the opposite effect, which suggests social immersion could have amplified social stimuli as (social) stressors, counteracting stress reduction. This mechanism is consistent with findings that training in intersubjective skills, compared with a mindfulness-based approach after exposure to psychosocial stressors, reduced the physiological stress response (measured by cortisol) (Engert et al., 2017). Another study (Faucher et al., 2016) suggests that cognitive behavioral therapy is slightly superior to the mindfulness-based approach for treating social anxiety disorder. These mixed findings underscore the need for further VR-aided mindfulness-based training research given the limited number of studies with substantial heterogeneity (Ma et al., 2023).

The randomized design, including both active and passive control groups, allowed for a rigorous comparison across intervention conditions. Moreover, implementing the intervention in a real-world university setting offers valuable insight into its ecological validity and practical applicability. However, the small sample size with a high proportion of female psychology students, limits generalizability. Specifically, our results may not fully extend to older populations, individuals with limited technological proficiency, or those with different cultural backgrounds. Methodological aspects, such as minor randomization issues (see SM1 in Supplementary material) and unmonitored compliance (exercise frequency and intensity), may also have impacted the results. Main effects over time, even in the passive control group, raise questions about true intervention effects vs. testing effects or reactivity. Furthermore, the lack of measures assessing subjective satisfaction with the at-home MBI may have introduced an unaccounted confounding variable. Therefore, future studies should increase the frequency of data collection (e.g., via weekly assessments or diary methods) to better capture intra-individual changes and dynamic, process-related situational factors (Pawsey et al., 2021), or physiological assessments (e.g., cortisol levels). The absence of follow-up assessments further limits insights into long-term effects. Future research should also explore variations of group meetings and VR environments, e.g., examining different VRs to clarify how they might enhance or hinder mental health training. While VR offers promising opportunities, its effectiveness may be influenced by individual differences in technology acceptance, susceptibility to motion sickness, and prior experience with immersive environments. Future research should, therefore, explore how these factors affect engagement and outcomes in more diverse populations. Along these lines, the integration of VR and artificial intelligence (AI) in mindfulness-based interventions is a promising research area (Mitsea et al., 2023). As technology advances, future studies should examine how these tools can be adapted for diverse populations and contexts, ensuring accessibility and reducing barriers (Eichel et al., 2021).

Our findings show that app-based mindfulness training, supported by digital group meetings, can be embedded into students' daily lives to strengthen stress management, reduce social isolation, and enhance self-compassion. These benefits translate into greater resilience when facing academic pressures, healthier responses to personal setbacks, and improved social connectedness. Video calls offer a practical and accessible format for regular peer support, while VR environments—when designed for comfort and usability—can provide immersive spaces that foster deeper focus and engagement. Looking ahead, refining VR platforms to minimize technical and physical barriers, and integrating adaptive features such as AI-driven personalization or real-time feedback, could further tailor mindfulness training to individual needs. Universities could harness these technologies to deliver flexible, scalable mental health support that evolves with students' learning contexts and lifestyles.