This study was a single-blind, parallel-group, superiority RCT evaluating an 8-week online self-help mindfulness-based stress management program as an adjunct to pharmacological treatment for patients with GAD. Participants were randomized at the individual level. The trial was prospectively registered at the Chinese Clinical Trial Registry (registration number: ChiCTR2300078470) on December 8, 2023, prior to enrollment of the first participant on December 20, 2023. This trial is reported in accordance with the CONSORT (Consolidated Standards of Reporting Trials) 2025 statement and the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) extension for web-based and mobile health interventions [38,39]. The completed CONSORT 2025 and CONSORT-EHEALTH checklists are provided as Checklist 1 and Checklist 2, respectively. This was a single-center trial conducted at Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China. Participants were consecutively recruited from the outpatient anxiety and depression clinics across 3 campuses of Tongji Hospital (Main Campus, Optical Valley Branch, and Sino-French New City Branch) between December 2023 and September 2024. Follow-up for the final participant, including outcome and adverse event assessment, was completed in December 2024. The full study protocol was not publicly posted or separately preregistered before trial completion, but it is available from the corresponding author upon reasonable request.

The study was approved by the Ethics Committee of Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology (approval number: TJ-IRB20230617) and was conducted in accordance with the Declaration of Helsinki and relevant ethical guidelines. In April 2024, an approved protocol amendment (TJ-IRB202404034) added electroencephalography and magnetic resonance imaging assessments for mechanistic exploration; these measures were not part of the original protocol and are not reported here. Written informed consent was obtained from all participants after they received a full explanation of the study purpose, procedures, and potential risks. Study data were deidentified prior to analysis and were accessible only to authorized research personnel. No financial compensation was provided for participation. The screenshot from the course interface shown in Figure S1 in Multimedia Appendix 1 was included with the written informed consent of the individual depicted. No other identifiable participant information is included in the manuscript or supplementary materials.

Participants were screened for eligibility according to the following inclusion and exclusion criteria. Inclusion criteria were as follows: (1) a primary diagnosis of GAD according to the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders [Fifth Edition]); (2) a Hamilton Anxiety Rating Scale (HAMA) score of ≥7; (3) age ≥18 years; (4) access to a smartphone or computer with a stable internet connection; and (5) sufficient cognitive and behavioral ability to follow the intervention procedures and complete study assessments. Exclusion criteria were as follows: (1) a diagnosis of schizophrenia, bipolar disorder, major depressive disorder with suicide risk, or substance use disorder (eg, alcohol or drug abuse) within the past 6 months; (2) receipt of formal psychotherapy (eg, CBT, psychodynamic therapy) or structured mindfulness training within the past 6 months; (3) severe visual, auditory, or cognitive impairments; (4) use of psychotropic medication (eg, antidepressants, antipsychotics, and benzodiazepines) within the past 3 months prior to enrollment; (5) pregnant, planned pregnancy, or breastfeeding; (6) concurrent participation in another clinical trial or enrollment of a first-degree relative in this study; (7) inability to cooperate in the intervention or data collection process; and (8) any other condition that, in the opinion of the investigators, would interfere with participation or data integrity. The primary diagnosis of GAD was established by the treating psychiatrist according to DSM-5 criteria during routine psychiatric evaluation. Inclusion criteria were used for eligibility screening. Major psychiatric exclusion criteria were assessed during the same evaluation based on clinical interview, clinical history, and available medical records. Detailed procedures are provided in Multimedia Appendix 1.

At baseline, all participants underwent a comprehensive assessment, including demographic information, clinical history, and psychological evaluations. Eligible participants were randomly assigned to either the internet-based stress management program (iSM) plus TAU group or the TAU group. The iSM+TAU group received internet-based mindfulness stress management combined with routine pharmacological treatment, while the TAU group received only pharmacological treatment. Assessments were conducted at baseline (T0), midintervention (T1, Week 4), postintervention (T2, Week 8), and at 3-month follow-up (T3) to evaluate efficacy and safety, including psychological assessments, medication adjustments, adherence, and adverse event monitoring.

Of 272 participants screened, 140 eligible participants were randomized to iSM+TAU (n=73) or TAU (n=67; Figure 1). Following randomization, 7 participants in the iSM+TAU group and 4 in the TAU group withdrew before initiating the intervention. At posttreatment, 59 participants in the iSM+TAU group and 49 in the TAU group completed the primary outcome assessment. By the 3-month follow-up, 54 participants in the iSM+TAU group and 41 in the TAU group had completed all assessments. Of the 73 participants randomized to the iSM+TAU group, 74% (54/73) completed at least 4 sessions, and the median number of sessions completed in the full intervention sample was 7 (IQR 3‐8). Missing data rates were 15.7% (22/140) at midintervention (4 wk, T1), 22.9% (32/140) at posttreatment (8 wk, T2), and 32.1% (45/140) at the 3-month follow-up (T3). Across all postbaseline assessments, 99 of 420 expected observations were missing, corresponding to an overall missingness rate of 23.6%. Little missing completely at random test was nonsignificant (χ²₁₇₂=184.2; P=.25), indicating no significant departure from missing at random.

No significant differences were observed in any baseline characteristics between the iSM+TAU and TAU groups (all P>.05; Table S2 in Multimedia Appendix 1). The median age of all participants was 33 (IQR 26‐39) years, and 66.4% (93/140) were female. The sample was predominantly urban and relatively well educated, 81.4% (114/140) were from urban areas, and 77.9% (109/140) had a college education or above. Medication use was similar between the 2 groups (P>.05), with selective serotonin reuptake inhibitor monotherapy being the most common regimen, accounting for 74% (54/73) in the iSM+TAU group and 70.1% (47/67) in the TAU group.

All randomized participants (n=140) were included in the primary ITT analyses. Linear mixed-effects models showed significant group×time interactions for both anxiety and depressive symptoms (P<.001). At posttreatment (T2), the iSM+TAU group demonstrated significantly greater reductions in anxiety symptoms (HAMA: mean difference [MD]=−1.994, 95% CI −3.758 to −0.230; P=.03; Cohen d=−0.277, 95% CI −0.521 to −0.033) and depressive symptoms (HAMD-21: MD=−1.649, 95% CI −3.165 to −0.133; P=.03; Cohen d=−0.309, 95% CI −0.592 to −0.026) compared with the TAU group. These between-group differences further increased at the 3-month follow-up, reaching moderate effect sizes for both HAMA (Cohen d=−0.511, 95% CI −0.755 to −0.267) and HAMD-21 (Cohen d=−0.637, 95% CI −0.920 to −0.355), as illustrated in Figures 2A and 2B. Detailed estimates are provided in Table S3 in Multimedia Appendix 1, and consistent findings were observed in the PP sensitivity analysis (Table S4 in Multimedia Appendix 1).

In the ITT sample, significant group×time interactions were found for PHQ-15, PSQI, S-AI, FFMQ, and PSS-10 (Table S3 in Multimedia Appendix 1). At posttreatment, the iSM+TAU group showed significant improvements in somatic symptoms (PHQ-15: MD=−1.388, 95% CI −2.465 to −0.312; P=.01; Cohen d=−0.340, 95% CI −0.604 to −0.077; Figure 2C), state anxiety (S-AI: MD=−6.254, 95% CI −9.905 to −2.602; P<.001; Cohen d=−0.537, 95% CI −0.849 to −0.224; Figure 2D), mindfulness level (FFMQ: MD=8.433, 95% CI 4.119-12.748; P<.001; Cohen d=0.666, 95% CI 0.327-1.006; Figure 2E), and perceived stress (PSS-10: MD=−2.801, 95% CI −4.740 to −0.863; P=.005; Cohen d=−0.429, 95% CI −0.725 to −0.133; Figure 2F). Sleep quality (PSQI: MD=−0.259, 95% CI −1.479 to 0.962; P=.68; Cohen d=−0.068, 95% CI −0.386 to 0.251) showed no significant between-group difference at posttreatment but demonstrated a significant effect at follow-up (MD=−2.243, 95% CI −3.464 to −1.023, P<.001; Cohen d=−0.588, 95% CI −0.906 to −0.269; Figure 2G).

Rumination (RRS: MD=−4.424, 95% CI −8.069 to −0.780; P=.02; Cohen d=−0.344, 95% CI −0.626 to −0.062) showed significant posttreatment reductions, although the overall group×time interaction was not significant (P=.38; Figure 2H). Social functioning measured by the SDSS showed no significant between-group difference at posttreatment (MD=−0.424, 95% CI −1.091 to 0.243; P=.21; Cohen d=−0.146, 95% CI −0.376 to 0.083), with significant effects emerging only at follow-up (MD=−0.947, 95% CI −1.614 to −0.279; P=.006; Cohen d=−0.327, 95% CI −0.557 to −0.097; Figure 2I). The PP sensitivity analysis revealed consistent patterns of change across all secondary outcomes (Table S4 in Multimedia Appendix 1).

No serious adverse events were reported during the trial. A limited number of participants (15/140, 10.7%) reported mild transient dizziness or nausea, which was considered more likely related to concurrent pharmacotherapy than to the internet-based behavioral intervention. No adverse events led to study withdrawal.

Of the 73 participants randomized to the iSM+TAU group, 7 (9.6%) did not complete a full session, 5 (6.8%) completed 1 session, 4 (5.5%) completed 2 sessions, 3 (4.1%) completed 3 sessions, 8 (11%) completed 4 sessions, 5 (6.8%) completed 5 sessions, 4 (5.5%) completed 6 sessions, 11 (15.1%) completed 7 sessions, and 26 (35.6%) completed all 8 sessions (Table S5 in Multimedia Appendix 1). The mean number of sessions completed in the full intervention sample was 5.3 (SD 2.9), and the median number of sessions completed was 7 (IQR 3‐8).

We then conducted exploratory analyses to examine whether greater intervention exposure was associated with better clinical outcomes. Among participants who completed all follow-up assessments, those who completed ≥4 sessions showed significantly greater anxiety reduction than those completing <4 sessions (HAMA: F_1,47=5.13; P=.03), but no significant difference was observed for depressive symptoms (HAMD-21: F_1,47=2.86; P=.10; Table 1). In addition, compared with the TAU group, participants in the iSM+TAU subgroup who did not complete the intervention showed no significant differences in either anxiety symptoms (HAMA: F_1,62=0.31; P=.58) or depressive symptoms (HAMD-21: F_1,62=0.01; P=.91; Table 1), suggesting that limited intervention exposure may have attenuated the potential treatment benefit.

Among iSM completers (n=54), several adjusted regression models were used to examine baseline predictors of anxiety reduction (HAMA change from baseline to posttreatment). After univariate screening (P<.05) and adjustment for covariates including baseline HAMA scores, age, sex, marital status, education level, residential status, employment status, and childhood parental separation, separate covariate-adjusted linear regression models showed that higher trait anxiety was significantly associated with smaller reductions in anxiety symptoms (standardized β=–0.150, 95% CI –0.234 to –0.002; P=.046), while greater acting with awareness predicted larger improvements (standardized β=0.167, 95% CI 0.031-0.335; P=.02; Table 2). These findings suggest that individuals with lower trait anxiety and higher baseline mindful awareness may be more suitable for self-guided digital mindfulness interventions.

Across the 3 intervals (T0-T1, T1-T2, and T2-T3), the temporal network structure revealed distinct dynamic patterns among the 6 retained variables: HAMA, HAMD-21, FFMQ, PSQI, S-AI, and PSS (Figure 3). In these temporal networks, nodes represent psychological variables, and directed edges indicate temporal predictive associations from time t to time t+1. Red edges denote positive associations, whereas blue edges denote negative associations. Solid lines indicate edges with at least 90% bootstrap support, whereas dashed lines indicate edges with less than 90% bootstrap support.

During the early treatment phase (T0-T1; Figure 3A), the strongest cross-lagged associations were observed between perceived stress and state anxiety, with PSS predicting subsequent S-AI (B=0.389, 95% CI 0.066-0.672) and S-AI, in turn, predicting PSS (B=0.351, 95% CI 0.084-0.570), suggesting a reciprocal temporal association between perceived stress and state anxiety. Higher baseline mindfulness (FFMQ) predicted lower subsequent perceived stress (B=−0.226, 95% CI −0.419 to −0.017), whereas higher state anxiety predicted reductions in mindfulness (B=−0.182, 95% CI −0.352 to −0.015). In addition, depressive symptoms predicted higher subsequent anxiety (HAMD → HAMA: B=0.237, 95% CI 0.058-0.501), and elevated state anxiety was associated with poorer subsequent sleep quality (S-AI → PSQI: B=0.164, 95% CI 0.093-0.323). In this interval, PSS showed the highest in-predictability, whereas S-AI exhibited the highest out-predictability (Figure S3 in Multimedia Appendix 1).

During midtreatment (T1-T2; Figure 3B), the PSS → S-AI pathway remained the strongest cross-lagged edge and further strengthened (B=0.442, 95% CI 0.214-0.621). Anxiety symptoms (HAMA) also showed a positive predictive effect on subsequent state anxiety (B=0.227, 95% CI 0.017-0.472). In parallel, poorer sleep quality predicted higher perceived stress (PSQI → PSS: B=0.205, 95% CI 0.003-0.408), while perceived stress predicted lower mindfulness (PSS → FFMQ: B=−0.151, 95% CI −0.322 to −0.044). State anxiety demonstrated the highest in- and out-predictability during this phase (Figure S3 in Multimedia Appendix 1).

During the consolidation phase (T2-T3; Figure 3C), the strongest cross-lagged association emerged between anxiety and depressive symptoms, with HAMA predicting higher subsequent HAMD-21 scores (B=0.311, 95% CI 0.086-0.481). Notably, mindfulness was negatively associated with subsequent state anxiety (B=−0.270, 95% CI −0.397 to −0.088), anxiety (B=−0.202, 95% CI −0.398 to −0.004), depression (B=−0.202, 95% CI −0.341 to −0.018), and perceived stress (B=−0.150, 95% CI −0.352 to −0.018). In this interval, depressive symptoms (HAMD-21) exhibited the highest in-predictability and out-predictability, suggesting that depressive symptoms were more strongly connected with subsequent symptom changes during the late phase (Figure S3 in Multimedia Appendix 1). Together, these findings suggest that higher mindfulness was associated with lower subsequent symptoms across multiple domains. Over time, the stress-anxiety pattern appeared to weaken, whereas the anxiety-depression linkage became more pronounced in the late phase.

The correlation stability coefficients for edge weights were 0.284 for T0-T1, 0.358 for T1-T2, and 0.358 for T2-T3, all exceeding the recommended minimum threshold of 0.25, indicating modest but acceptable stability of edge estimates across intervals. In contrast, correlation stability coefficients for in-strength and out-strength centrality were below 0.25 in most intervals, and these centrality metrics were therefore not interpreted.

To reduce the risk of spurious causal inferences arising from stable between-person differences in conventional cross-lagged models, RI-CLPM analyses were conducted to isolate within-person lagged associations. Random intercepts capture stable between-person differences (trait-like components) across time. Squares and uppercase letters indicate observed variables, whereas ovals and lowercase letters denote latent within-person components. Autoregressive (α) paths represent temporal stability within constructs, and cross-lagged (γ) paths represent directional within-person effects between constructs across adjacent time points. Residual terms (u_₁-u_₃ and v_₁-v_₃) represent time-specific unexplained within-person variance. All autoregressive and cross-lagged paths are constrained to be equal across time points, estimating lagged effects from time t to time t+1 across 4 measurement waves. Only 2 pathways showed significant within-person effects: the bidirectional pathway between PSS and S-AI (Figure 3D) and the unidirectional pathway from FFMQ to S-AI (Figure 3E). Both models demonstrated good fit with appropriate variance partitioning between stable traits and temporal fluctuations (see Multimedia Appendix 1).

Cross-lagged paths revealed robust bidirectional within-person associations (Figure 3D): greater perceived stress predicted higher subsequent state anxiety (b=0.363, 95% CI 0.163-0.563, P<.001; mean β=0.219), and higher state anxiety predicted greater subsequent stress (b=0.062, 95% CI 0.012-0.111, P=.02; mean β=0.165). Both effects exceeded the large effect size threshold (β≥0.12), supporting a reciprocal reinforcement mechanism at the individual level (Table 3).

A significant unidirectional within-person effect was observed (Figure 3E): higher mindfulness predicted lower subsequent state anxiety (b=−0.210, 95% CI −0.326 to −0.093; P<.001; mean β=−0.285), whereas state anxiety did not significantly predict later mindfulness (b=−0.067, 95% CI −0.164 to 0.031; P=.18). These findings suggest that increases in mindfulness precede reductions in state anxiety over time (Table 3).

This study evaluated a brief, self-guided, internet-delivered mindfulness-informed stress management program as an adjunct to TAU for patients with GAD. Overall, the intervention was associated with improvement in anxiety, depression, and several functional outcomes, while also showing acceptable adherence and feasibility in a real-world clinical setting. In addition, the longitudinal network-based analyses provided exploratory evidence suggesting that stress- and mindfulness-related symptom processes may be involved in how improvement unfolded over time.

The present study demonstrated significant anxiety reduction in the iSM+TAU group compared with TAU, with effects that were sustained and larger at the 3-month follow-up. This pattern of sustained and growing improvement suggests that skills acquired during training continue to consolidate after formal intervention ends [57]. Although the effect size was lower than that found in a comparable RCT with more intensive interventions in a therapist-guided online mindfulness intervention [55], it was larger than the small effects reported in a meta-analysis of online mindfulness-based interventions [22] and consistent with recent work suggesting that fully self-guided digital mindfulness approaches can be acceptable and beneficial for adults with GAD [58].

Beyond symptom reduction in anxiety and depression, iSM demonstrated broad functional improvements, including sleep quality, somatic symptoms, and social functioning. These multidimensional benefits are particularly important in real-world clinical settings where functional recovery often lags behind symptom improvement [59-62]. More broadly, mindfulness-based interventions may have effects across multiple domains rather than on emotional symptoms alone, with meta-analytic evidence suggesting benefits for global cognition and several cognitive control–related processes [63]. Although these domains were not directly examined in the present study, they may provide a broader framework for understanding symptom improvement. Notably, perceived stress, a central focus of the intervention, was significantly reduced by midtreatment, suggesting that stress reduction may be an early and clinically relevant component of improvement.

The retention rate of 77% (108/140) and median completion of 7 sessions demonstrate excellent feasibility for a self-directed intervention. Participant retention in the present trial was comparable to that reported by Li et al [64] and notably higher than rates documented in other large-scale RCTs of online mindfulness-based interventions, where attrition has reached approximately 38.7% [65]. Importantly, most dropouts occurred before or during the initial sessions, suggesting that early engagement is critical for retention [66]. The relatively high adherence to this self-guided format underscores the feasibility of iSM, particularly in settings with limited therapeutic resources [20,21]. Treatment adherence was closely associated with outcomes; participants who completed more sessions showed greater anxiety reduction, whereas those with incomplete participation derived little benefit beyond TAU. Dose-response analyses further indicated that completing at least 4 sessions, covering core mindfulness components, was associated with greater improvement and may represent a pragmatic minimum exposure target in self-guided settings [67].

Individual psychological characteristics appeared to significantly influence the effectiveness of the intervention. Higher baseline acting with awareness—a mindfulness facet reflecting the tendency to be attentive to present-moment activities—predicted greater improvements, consistent with previous research suggesting that existing mindfulness capacity facilitates stress management skill acquisition [68]. This may reflect better ability to engage with training exercises or greater openness to experiential approaches. Conversely, elevated trait anxiety was associated with attenuated benefits, potentially reflecting impaired attentional control that limits effective engagement with stress management techniques [69]. Although these predictor effects were modest, they may still be informative as partial contributors to treatment response variability. From a clinical perspective, these findings suggest that individuals with higher baseline mindfulness capacity may be well-suited for self-guided digital programs as adjuncts to standard care, whereas those with very elevated trait anxiety may require more intensive or therapist-supported interventions. Such stratification could maximize both clinical outcomes and health care efficiency.

While traditional RCT analyses established the overall efficacy of the intervention, network analyses were used to explore how symptom associations changed over time and to generate hypotheses about possible temporal processes underlying change. CLPN analyses suggested stage-dependent changes in the pattern of temporal symptom associations across the study period. In the initial phase, perceived stress and anxiety showed a relatively prominent bidirectional association, a pattern broadly consistent with network theories emphasizing persistent interconnections among symptoms despite reductions in overall symptom levels [70]. This pattern appeared to weaken over time, whereas anxiety remained highly connected and stress and sleep continued to show temporal associations with other variables. In later phases, depressive symptoms appeared more strongly embedded in the temporal network and were more often linked to subsequent changes in other variables. Meanwhile, higher mindfulness was generally associated with lower subsequent levels of stress, anxiety, and depression across phases. These findings are broadly consistent with nonlinear models of psychological change, which propose that symptom improvement may occur through stage-dependent reorganization [71].

However, not all temporal associations identified in the CLPN analyses remained evident under the RI-CLPM framework. RI-CLPM, which more explicitly accounts for stable trait-like between-person differences, suggested that 2 pathways remained relatively robust under this more stringent model and were broadly compatible with within-person temporal dynamics. First, the stress-anxiety relationship was bidirectional: perceived stress predicted subsequent anxiety, and anxiety predicted subsequent stress, a pattern broadly consistent with reciprocal processes described in transactional stress models [72]. Second, mindfulness showed a unidirectional protective effect on anxiety, whereas the reverse path was not supported in the model, suggesting that mindfulness may function as a protective regulatory process, although this interpretation remains inferential and model-based, consistent with attentional control theory [73]. Other associations identified in the CLPN analyses were not retained in the RI-CLPM framework, suggesting that they may be less robust as within-person temporal pathways. Taken together, these findings are broadly consistent with theoretical models emphasizing stress reactivity, attentional regulation, and stage-dependent symptom change and suggest that mindfulness-related processes may contribute to symptom improvement over time [31,72,73].

The present study provides new evidence for a brief, self-guided, culturally adapted internet-based mindfulness-informed stress management program evaluated as a feasible and clinically useful adjunct to pharmacological treatment for GAD in a Chinese clinical setting. Unlike prior work focused mainly on symptom outcomes alone, it also incorporated exploratory longitudinal analyses to provide a more process-oriented perspective on how symptoms may change over time, suggesting that perceived stress and mindfulness may represent potential processes involved in treatment response and warrant further investigation. The iSM may also help expand access to mental health care in settings with limited psychotherapy resources. As a brief digital program requiring only basic internet access and no therapist scheduling, iSM may offer a practical and relatively low-cost option for wider implementation, especially within stepped-care models and among underserved populations [74,75]. The integration of culturally relevant elements further enhances engagement and acceptability, supporting iSM’s potential as an equitable adjunct to traditional services within evolving mental health systems.

This study has several limitations. First, although participants were recruited across 3 hospital campuses, this study was conducted within a single hospital system with a relatively homogeneous sample, limiting generalizability. The sample profile may partly reflect self-selection into a self-guided digital intervention, which may be more readily adopted by younger and more educated participants. These factors highlight the need for future adaptations and testing in broader populations. In addition, the achieved sample size was slightly below the pre-registered recruitment target due to funding and time constraints. Future studies should expand the sample size and include multiple centers with more diverse populations across different regions, while also considering adaptations that may improve accessibility for broader clinical groups and enhance external validity. Additionally, the CLPN and RI-CLPM analyses were exploratory and should be interpreted cautiously. Future studies with larger samples are needed to verify the robustness and reproducibility of these longitudinal findings. Second, the follow-up period was limited to 3 months after the intervention, making it difficult to evaluate long-term effectiveness and adherence. Future research should extend the observation period and incorporate strategies such as booster sessions and individualized guidance to support sustained engagement and outcomes. Third, this study primarily relied on standardized scales for assessment, which may be subject to expectancy effects. Future research should incorporate objective physiological and neuroimaging measures to improve the precision of mechanism evaluation. Fourth, the network analysis was restricted to 6 theory-driven variables capturing the hypothesized stress-management pathway and key GAD-related symptom domains, which helped preserve model interpretability and stability given the sample size and repeated-measures design, but other potentially relevant mechanisms (eg, emotion regulation and self-efficacy) were not examined. Future studies with larger samples and broader assessments should investigate more comprehensive mechanistic networks. Fifth, the use of TAU rather than an active control also limits the ability to distinguish intervention-specific effects from nonspecific influences such as attention, expectancy, or engagement with a digital platform, although this design was consistent with the pragmatic aim of evaluating added benefit beyond usual care. Finally, while RI-CLPM strengthens within-person temporal inference by controlling for stable individual differences, causal conclusions still require experimental manipulation of specific mechanisms.

To our knowledge, this is the first RCT to evaluate a brief, self-guided, culturally adapted digital mindfulness intervention as an adjunct to pharmacotherapy in Chinese adults with GAD. The intervention yielded clinically meaningful benefits across anxiety, depression, and several secondary outcomes related to stress, mindfulness, and associated clinical domains. Its fully self-guided format, culturally adapted content, and minimal infrastructure requirements highlight its practical potential as a scalable, resource-efficient approach in real-world settings where psychotherapy availability is limited. Unlike prior studies focused mainly on symptom outcomes alone, this trial combined a randomized design with exploratory CLPN and RI-CLPM analyses to provide preliminary insight into symptom change processes over time. The longitudinal findings suggested temporally distinct dynamics, with stress-anxiety coupling more prominent early in treatment and mindfulness-related pathways becoming more apparent in later phases. These findings contribute not only preliminary evidence of adjunctive clinical utility but also add a more process-oriented analytic perspective that may inform future intervention refinement.