Data collection for the trial took place between November 21, 2021, and April 2023. Participants with a current depressive episode were recruited from “Merhavim” Medical Center. Eligible participants from all racial, ethnic, and gender categories were aged between 18 and 65 years. A minimum score of 13 on the 17-item Hamilton Rating Scale for Depression (HDRS-17) [56] was required. Additional inclusion criteria were a primary DSM-5 (Diagnostic and Statistical Manual of Mental Disorders [Fifth Edition]), diagnosis of MDD, with a single or recurrent episode, with the additional requirement of a current episode with a duration of ≥4 weeks. Participants needed to be without antidepressant medication or on an adequate (ie, at least the lowest effective) and stable (ie, for ≥6 weeks) dosage. Concomitant psychotherapy was allowed if ongoing for at least 3 months prior to baseline. Participants were excluded from the study if they met any of the following criteria: problematic gaming, as assessed by the Gaming Disorder Test [57]; a diagnosis of a significant neurological disorder or comorbid psychiatric disorder; any relevant unstable medical condition (eg, acute or unstable cardiac disease); use of an investigational drug within 6 weeks prior to baseline; or suicidal risk, indicated by a score of 4 to 6 on item 10 of the MADRS or a suicide attempt during the current depressive episode.

Participants were recruited through clinical referrals from psychiatrists, psychologists, and social workers at Merhavim Medical Center. Interested individuals were contacted via phone by a member of the study team and invited to an in-person screening visit (V₀) at Merhavim, where eligibility was assessed. No flyers, social media, or website-based recruitment were used.

A power analysis was conducted using G*Power (version 3.1.9.7; Heinrich Heine University Düsseldorf) [58]. Sample size estimation was based on the CC outcome. Prior randomized trials of computerized cognitive training in individuals with elevated depressive symptoms have reported medium-to-large effects on clinical outcomes (d≈0.73 on depressive symptoms [24]). This effect size assumption was further supported by meta-analytic estimates of computerized cognitive training effects on depressive symptoms [30] and by earlier cognitive-emotional training trials in MDD and in older adults [32,33,39,40]. To allow for a more conservative estimate, we planned for a medium effect size (Cohen d=0.65). Under this assumption, a total sample size of 60 participants (30 per group) provides 80% power to detect a between-group difference at an α of .05. These effect size estimates were derived largely from trials using inactive or minimal-treatment comparators. A comparison against a closely matched active control likely corresponds to a smaller true between-arm difference; the assumed effect size of 0.65 should therefore be regarded as optimistic, and the trial was likely underpowered for the effect sizes expected between 2 engaging digital interventions (see “Limitations” section).

The study was a double-blind randomized controlled trial. Figure 1 depicts the study design. Interested participants were screened for eligibility (V₀). Eligible participants were assessed at baseline (V₁) and were then randomized in a 1:1 ratio to receive either the active LoH game intervention or the matched active control arm (LoH_control).

Randomization used a validated computer-generated pseudorandom sequence with permuted blocks of 4 (1:1 allocation ratio), prepared by the co-principal investigator (PI), who had no contact with participants. Participants were assigned sequentially in order of enrollment. The allocation list was held by the PI; after completion of baseline assessment, the PI released the participant’s assignment to the unblinded staff member responsible for intervention distribution. Outcome assessors remained blinded to allocation. No additional allocation concealment procedures (eg, centralized automated assignment) were used. Both the active and the control games were self-administered, and study staff were available to assist remotely whenever needed.

The trial was prospectively registered with depressive symptom change on the MADRS as the primary end point, feasibility as the secondary end point, and tertiary end points spanning symptoms, well-being, rumination, function, and participation. Exploratory end points included CC change, durability across the 1-month follow-up, as well as treatment expectancy and intrinsic motivation. This analysis follows this structure; intrinsic motivation is reported within the secondary feasibility framework—consistent with standard practice in feasibility studies of digital mental health interventions (deviation is disclosed here per CONSORT recommendations). A within-arm exploratory analysis testing whether CC gains are associated with depressive symptom reductions is reported as an additional mechanistic analysis (not formally preregistered as a hypothesis).

The trial was designed as a feasibility and effectiveness study, and its outcome hierarchy follows the preregistration and was not altered post hoc. The MADRS was chosen as the primary (effectiveness) end point following clinical standards for depression trials. Feasibility was the secondary end point because, given the novelty of this home-based, minimal-contact serious game intervention, establishing that the protocol could be delivered and completed per protocol was a central goal of this first trial. CC was exploratory, assessed to probe the proposed training mechanism rather than as a confirmatory clinical outcome, and, because the assessment battery shares task demands with the LoH training mini-games, change on it cannot be cleanly interpreted as generalizable transfer (see “Discussion” section).

The preregistered primary outcome was change in depressive symptoms on the MADRS [59], from baseline (V₁) to postintervention (V₄), reflecting the hypothesis that the LoH intervention would reduce depressive symptoms more than the active control game. Statistical inferences for the primary outcome were performed at an α of .05. The MADRS is a clinician-administered scale with 10 items scored on a 7-point scale (0‐6). The total score ranges between 0 and 60, with higher scores indicating greater depression severity.

Feasibility outcomes comprised (1) adherence (completed training sessions and total minutes played); (2) retention at postintervention (V₄) and follow-up (V₅); and (3) intrinsic motivation, assessed at the end of week 1 and at V4 with the Intrinsic Motivation Inventory (IMI [60]). All 5 IMI subscales (interest/enjoyment, effort/importance, perceived competence, perceived choice, and pressure/tension) are reported separately; pressure/tension serves as a proxy for subjective tolerability (lower scores indicate greater tolerability).

The preregistered feasibility threshold was completion of ≥20 of the 30 prescribed sessions by ≥50% of randomized participants. We additionally applied 3 post hoc benchmarks drawn from comparable digital mental health trials [40,61,62]: retention at V₄ ≥70%, mean adherence ≥60%, and mean IMI interest/enjoyment in the active arm above the scale point (4/7)—framed as reference points rather than prespecified criteria. Expanded session-level in-game involvement metrics from server-side gameplay logs are reported in Multimedia Appendix 2; these correspond to the “training progress” pillar of the 4-pillar cognitive training evaluation framework of Harvey et al [63] (relabeled from “training engagement” in the framework of Harvey et al to avoid confusion with the engagement framework of Plass et al [64]).

Tertiary outcomes were chosen because CC deficits in MDD are theorized to operate through perseverative negative thinking that, in turn, impairs everyday functioning and reduces well-being. The preregistered selection also follows current state-of-the-art recommendations for computerized cognitive training trials [63]. It included MADRS-based response (≥50% reduction from baseline) and remission (MADRS ≤10) rates. The other tertiary outcomes capture (1) near-transfer—depressive symptoms; (2) far transfer—severity of illness, global improvement, and rumination; and (3) ecological transfer—well-being, functional capacity, plus participation in daily activities. Depression severity, a near-transfer measure, was assessed using the Patient Health Questionnaire-9 (PHQ-9 [65]), a self-reported measure with scores ranging from 0 to 27. Severity of illness, a more distant transfer measure, was assessed using the Clinical Global Impressions-Severity (CGI-S) scale [66]. The CGI-S is a clinician-rated measure used to evaluate the overall severity of a patient’s illness at a given time based on the rater’s clinical experience, ranging from 1 (normal, not at all ill) to 7 (among the most extremely ill patients). Global improvement was assessed using the companion Clinical Global Impressions-Improvement (CGI-I) scale [66], a clinician-rated measure of change relative to the patient’s condition at admission, ranging from 1 (very much improved) through 4 (no change) to 7 (very much worse). Rumination was assessed using the full Ruminative Responses Scale (RRS-22 [67]), a 22-item self-reported scale measuring participants’ tendency toward repetitive negative thinking. Items are rated on a 4-point Likert scale, ranging from 1 (almost never) to 4 (almost always), and a total score is derived, with higher scores indicating greater levels of rumination.

Concerning ecological transfer, well-being was measured using the World Health Organization (WHO)-5 Well-Being Index [68], a 5-item self-reported scale with positively worded statements scored on a 6-point Likert scale, ranging from 0 (none of the time) to 5 (all of the time), assessing subjective well-being over the past 2 weeks. Higher scores reflect better well-being. Functional capacity was assessed using the University of California, San Diego, Performance-based Skills Assessment (UPSA [69]), which evaluates functional capacity in 5 domains: medication management, financial management, using the telephone and communication, using public transportation, and planning leisure activities, using a simulation of 11 daily life tasks. Final scores range from 0 to 100. The Adult Subjective Assessment of Participation (ASAP [70]), a self-reported questionnaire, documents subjective and objective participation dimensions in daily life occupations of adults. Indices of objective participation include diversity (the percentage of activities in which the person participates) and intensity or frequency (ie, the prevalence of participation in activities). This measure also tracks activities that were stopped. Subjective dimensions include enjoyment (perceived pleasure in activities) and satisfaction (perceived fulfillment from participation), both rated on 7-point scales.

Change in CC was preregistered as an exploratory outcome, motivated by the mechanistic hypothesis that the LoH intervention trains CC. It is reported alongside the primary and tertiary end points in the “Results” section. CC was assessed using a computerized cognitive assessment battery, the Adaptive Cognitive Evaluation (ACE; Neuroscape, University of California, San Francisco [71]). We administered 6 ACE tasks, collectively measuring inhibitory control, cognitive flexibility, WM updating, sustained attention, and visuospatial WM. One of the tasks (Fileter, measuring visual WM) was excluded from analyses because many participants did not understand the task demands. On the basis of factor analyses of the remaining 5 tasks, a CC composite score was created from normed scores. ACE task outcomes and factor analysis are given in Multimedia Appendix 3.

Additional exploratory analyses included treatment credibility and outcome expectancy, assessed at the end of week 1 and at V₄ with the Credibility/Expectancy Questionnaire (CEQ [72]; credibility 3‐27, expectancy 3‐27, higher scores indicate greater credibility/expected benefit). In addition, durability of treatment effects was defined as change in all outcomes between baseline (V₁) and 1-month follow-up (V₅) and is reported for each end point alongside the V1 to V4 analysis.

The intervention is described using the TIDieR (Template for Intervention Description and Replication) checklist categories; a TIDieR summary table covering both arms is provided in Multimedia Appendix 4. The interventions were developed using Unity (version 2019.4.28f1; Unity Technologies). LoH and LoH_control versions 1.1.10 to 1.1.13 were used, corresponding to early bug corrections in the overall running performance, save system, and data collection, none of which affected the game content per se.

The distinction between the experimental (LoH) and active control (LoH_control) arms is not between a cognitively demanding and a cognitively inert condition, but between involvement matched and mechanism specific. Both arms load on cognitive engagement in the sense of Plass et al [64], offering structured goals, unambiguous feedback, and clear challenges. Critically, LoH_control was designed to match LoH on the patient’s ability to become involved with the game and thus spend time on task (involvement matched) while omitting the action mechanics hypothesized to challenge and train attention and CC (mechanism specific). In addition, LoH leverages both cognitive and affective engagement [64]—the latter via emotionally valenced facial expressions in the N-back and Go/No-Go modules, which target the mood-congruent attentional bias implicated in MDD [30-33]—whereas LoH_control relies exclusively on cognitive engagement. Importantly, LoH_control was designed to have a similar look and feel as the active LoH game and to also induce robust in-game involvement.

Interventions were developed in Unity and distributed directly to participants through tablets with the assigned game installed on them. Gameplay metrics (session duration and in-game involvement metrics) were automatically logged and transmitted to the study server. Each game was played during 30-minute sessions, within which several tasks were iterated. Both game versions place the player in an adventure world with tasks and subtasks while they discover the Village, a hub designed to maintain the players’ engagement by allowing them to exchange the points earned during play to personalize their game world via various aesthetic upgrades. Participants in both arms received identical instructions regarding gameplay and schedule.

The game LoH is designed as a cognitive training platform that integrates emotional and cognitive elements through various interactive challenges. The game architecture combines fast-paced action video game mechanics—centered around a core action game hub where players control a friendly creature navigating flying sequences that demand both divided and focused attention [73]—with targeted cognitive training mini-games. LoH is specifically designed to provide structured variability and adapt to individual player abilities through a multidimensional, probabilistic dynamic difficulty adjustment [74]. By continuously adapting task demands across multiple parameters, this system maintains each player within their zone of proximal development throughout training [74], sustaining the level of controlled processing required for cognitive skill acquisition—beyond merely reducing boredom or frustration [51].

LoH includes 7 such dynamically adjusted mini-games, each targeting a specific CC ability such as WM, inhibitory control, cognitive flexibility, sustained attention, divided attention, and multitasking, as detailed in Multimedia Appendix 4. Additionally, a deductive reasoning task is included as a supplementary, nonadaptive mini-game. LoH incorporates emotional stimuli (eg, emotionally valenced facial expressions in the N-back and Go/No-Go tasks) to specifically target the mood-congruent attentional bias implicated in MDD—that is, the tendency to direct attention toward negatively valenced stimuli [7]. Positive reinforcement elements (collecting coins, personalized rewards) provide additional motivational scaffolding [75]. An earlier version has been shown to successfully enhance attentional control in children [76].

The active control game (LoH_control) was designed to match the aesthetic and structural elements of the experimental game, an important feature to maintain experimental blinding. Critically, its mini-games lack the adaptive cognitive training components of LoH and its action-based high attentional demands. It consists of a central hub game and 4 distinct mini-games (a modified, no-action version of the flight hub, visual search, peripheral vision task, and decoration). These activities emphasize the slow and deliberate movement of focused attention through tasks such as guiding a bird character to collect resources, focusing on auditory cues in a peripheral vision task, or spotting concealed targets in a cluttered map. These activities were selected based on prior research that used similar tasks as control conditions for depressed or dysphoric participants [26]. As such, the control game does not challenge attentional control as much as LoH. Moreover, unlike LoH, which incorporates positive emotional reinforcement through collecting happy coins and working with happy facial expressions, the control game lacks emotionally salient features. Players were presented with a fixed sequence of these 4 mini-games, each played twice within the same training session. A description of the control mini-games is given in Multimedia Appendix 4.

All statistical analyses were conducted using R (R Foundation for Statistical Computing) [77,78]. Analysis was conducted by a person blinded to the group assignment (Mattan S Ben Shachar). Baseline measures were compared between groups, and the association between attrition and baseline measures was assessed, using t tests and chi-square tests. The primary analysis followed the intention-to-treat principle, retaining all randomized participants in their originally assigned groups. All outcomes with a baseline assessment were analyzed using linear mixed models (LMMs). LMMs with maximum likelihood estimation provide unbiased estimates under the missing-at-random assumption and were used as the primary handling strategy. We report here ANOVA-like omnibus F tests with Satterthwaite degrees of freedom. These models include fixed effects for time (baseline, postintervention, and follow-up), group (active vs control), and the time×group interaction. Random intercepts were specified for participants to account for individual variability at baseline.

As loss to follow-up at postintervention was strongly associated with the number of completed training sessions rather than with baseline symptom severity, the number of completed sessions (mean centered) was included as an additional covariate in all outcome models to reduce potential attrition-related bias and to make the missing-at-random assumption more plausible. This covariate was not part of the preregistered analysis plan and is itself a postrandomization variable; we therefore also report unadjusted models (omitting this covariate) as a sensitivity analysis in Multimedia Appendix 5, which yielded materially identical conclusions.

For significant time×group interactions, post hoc comparisons with Bonferroni correction were conducted. Effect sizes are reported using partial η²_p with 2-tailed 90% CIs, and estimated mean differences are reported with 95% CIs. Statistical inference for the primary outcome (H1: MADRS group×time) was performed at an α of .05. Tertiary outcome (H3) and exploratory analyses (CC, durability, mechanism, and predictors) are reported with unadjusted P values.

The secondary feasibility outcomes include the preregistered feasibility threshold (H2: ≥50% of randomized participants completing ≥20 of the 30 prescribed sessions) assessed as a descriptive proportion against the registered threshold. Feasibility, in-game involvement, and intrinsic motivation metrics were evaluated using descriptive statistics, independent-sample t tests for continuous measures, and chi-square tests (with Yates continuity correction) for categorical comparisons. Session-level gameplay-log comparisons (early abandonment and per-session active gameplay duration) were analyzed with mixed effects models, including a random intercept for participant to account for the nesting of sessions within participants (Multimedia Appendix 2). As an exploratory analysis, we examined the association between within-participant changes in CC and change in depressive symptoms. We first computed residual scores by regressing baseline scores out of postintervention and follow-up measurements separately. This approach allowed for the examination of change independent of initial differences, controlling for baseline variability and regression to the mean.

Participant-level missing data (loss to follow-up) were 16% (9/57) at postintervention (V₄) and 23% (13/57) at 1-month follow-up (V₅). Item-level missing data within completed assessments were minimal. The primary analytic strategy of linear mixed models with maximum likelihood estimation under the missing-at-random assumption (described earlier) accommodates this level of missingness. We assessed the missing data mechanism using Little test of missing completely at random; the test was nonsignificant for the primary outcome and for the other clinical measures, with a single borderline exception. As a sensitivity analysis, we repeated the primary group×time analyses under multiple imputation by chained equations (m=20). Both analyses are reported in Multimedia Appendix 6 and yielded conclusions consistent with the primary models.

Full feasibility metrics are provided in Multimedia Appendix 2. The preregistered feasibility threshold for gameplay completion was met: 30 of 57 (52.6%; 95% CI 39.9%‐65%) randomized participants completed ≥20 of the 30 prescribed sessions, exceeding the preregistered ≥50% threshold; among the 53 participants with recoverable gameplay data, 30 of 53 (56.6%; 95% CI 43.3%‐69%) met this threshold.

The 3 additional post hoc benchmarks were also met. First, retention was high at both postintervention (48/57, 84%; ≥70% benchmark) and follow-up (V₅: 44/57, 77%). Attrition was comparable across arms (active: 6/28, 21%; and control: 7/29, 24%; Figure 2). Specifically, 5 participants in the control arm discontinued participation during the first 2 weeks of the intervention, and 2 discontinued training later but completed the postintervention (V₄) assessment before being lost to follow-up. In the active arm, 2 participants discontinued their participation during the first 2 weeks of the intervention and 2 after week 2 (V₂); 2 further participants discontinued training later but completed the V4 assessment before being lost to follow-up. The reasons for discontinuation are reported by arm in Figure 2; the most common were difficulty meeting the protocol’s time requirements and finding the intervention unsuitable or not beneficial. Second, the mean adherence exceeded the 60% benchmark—participants completed, on average, 20.7 (SD 9.48) sessions of the 30 allotted sessions (mean 69%, SD 31.6%), with no significant between-group difference (t₅₀=1.25; P>.05). Third, the intrinsic motivation benchmark was met, with mean IMI interest/enjoyment in the active arm of 4.56 (SD 1.49) exceeding the scale midpoint (4/7).

Server-side gameplay logs were recovered for 53 of 57 randomized participants (LoH: 27/28, 96%; and LoH_control: 26/29, 90%) due to hardware failure. The mean active gameplay duration per session was 26.2 (SD 8.7) minutes for LoH and 28.7 (SD 8.3) minutes for LoH_control, closely matching the prescribed 30-minute dose. Early abandonment (active gameplay <15 minutes) was uncommon and did not differ between arms (LoH: 41/503, 8.2% vs LoH_control: 30/564, 5.3%; mixed-effects logistic regression with a random intercept for participant, z=1.76; P=.08). Sessions were distributed across all day parts, and 25% occurred on weekends (LoH: 126/503, 25%; LoH_control: 137/564, 24.3%), with similar patterns in both arms, indicating integration into varied daily schedules. Weekly mean session duration was stable across weeks 1 to 6 in both arms, with no evidence of gameplay fatigue. The catch-up provision at weekends and/or in week 7 was used by 27 of 27 LoH and 25 of 26 LoH_control participants and accounted for approximately 28% of all sessions in both arms (LoH: 143/503, 28.4%; LoH_control: 157/564, 27.8%). Per-participant adherence was heterogeneous: 33% (9/27) of LoH and 58% (15/26) of LoH_control participants completed ≥80% of prescribed sessions, while 30% (8/27) of LoH and 23% (6/26) of LoH_control completed fewer than half. This range is typical of remote digital mental health interventions and represents a target for personalization in future trials. Among discontinuers, all completed at least two sessions before quitting. Full distributions, statistical tests, and figures are reported in Multimedia Appendix 2.

Full results are given in Multimedia Appendix 8. At the end of week 1, IMI ratings were broadly comparable between arms for all subscales, except for perceived competence, which was higher in the control arm (t_39.98=−2.74; adjusted P=.03; mean difference −0.90, 95% CI −1.56 to −0.24).

At postintervention (V₄), the active arm reported higher interest/enjoyment (t_47.79=2.71; adjusted P=.02; mean difference 1.15, 95% CI 0.30-2.00) and higher effort/importance ratings (t₄₈=2.84; adjusted P=.02; mean difference 0.93, 95% CI 0.27-1.59) than the control arm, indicating that the active intervention was perceived as more enjoyable and more effortful. Pressure/tension scores were in the low-to-moderate range in both arms at both time points (LoH: mean 3.62, SD 1.58 at week 1; mean 3.90, SD 1.37 at V₄; LoH_control: mean 3.08, SD 1.21 at week 1; mean 3.13, SD 1.56 at V₄). Between-group comparisons were nonsignificant at week 1 (t_45.0=1.38; P=.17, mean difference 0.54, 95% CI −0.25 to 1.33) and at V4 (t_47.9=1.85; P=.07; d=0.52, mean difference 0.77, 95% CI −0.06 to 1.60). Although the V4 difference was not statistically significant, the effect size estimate (d=0.52) was in the direction of higher pressure/tension in LoH, consistent with the more attentionally demanding action game format.

MADRS response (≥50% reduction from baseline) and remission (MADRS ≤10) rates at postintervention and 1-month follow-up were computed, comparing groups using 2-sided Fisher exact tests with nonresponder imputation for missing MADRS scores. The results, indicating no between-group differences, are summarized in Multimedia Appendix 9.

Changes from V₁ to V₄ showed a significant effect of time (F_1,49.16=10.72; P=.002; η²_p=0.18, 90% CI 0.04-0.33). However, there were no significant effects of group (F_1,49.8=0.18; P=.67; η²_p=0.004) nor a significant group×time interaction (F_1,49.8=0.08; P=.78; η²_p=0.002; Figure 3A,B).

The effect of time was not statistically significant (F_1,98=3.75; P=.06; η²_p=0.04, 90% CI 0.00-0.11). However, there was no significant effect of group (F_1,98=0.97; P=.33; η²_p=0.01), nor a group×time interaction (F_1,98=0.43; P=.51; η²_p=0.004).

World Health Organization-Five Well-Being Index (WHO-5) demonstrated a significant improvement over time (F_1,46.5=5.17; P=.03; η²_p=0.10, 90% CI 0.01-0.24). However, there were no significant effects for group (F_1,46.6=0.27; P=.60; η²_p=0.006) nor a group×time interaction (F_1,47.03=0.06; P=.80; η²_p=0.001).

Severity of illness, measured using the CGI-S scale, demonstrated a significant improvement over time (F_1,49.2=13.55; P<.001; η²_p=0.22, 90% CI 0.07-0.36). However, there were no significant effects for group (F_1,52.8=0.003; P=.96; η²_p=0.00) nor a group×time interaction (F_1,50.1=0.52; P=.47; η²_p=0.01). Post hoc within-group comparisons showed a nonsignificant reduction from baseline to postintervention in the active arm (−0.57 points; 95% CI −1.11 to −0.03; t_48.8=−2.10; Bonferroni-adjusted P=.08) and a significant reduction in the control arm (−0.84 points; 95% CI −1.38 to −0.30; t_48.3=−3.11; Bonferroni-adjusted P=.006). As these P values are Bonferroni-adjusted for multiple comparisons and the CIs are unadjusted, the active-arm interval can exclude zero while its adjusted P value exceeds .05.

Global improvement was assessed with the clinician-rated CGI-I scale, which rates change relative to admission (1=very much improved to 7=very much worse; 4=no change). As CGI-I is defined relative to admission and therefore has no baseline value, it was not entered into the baseline-to-follow-up mixed models used for the other measures; instead, improvement was summarized descriptively and between-arm differences were tested at each postbaseline assessment. Participants were rated as improved: pooled CGI-I scores were significantly below the “no change” anchor at both postintervention (mean 2.96, SD 1.40; t₄₇=−5.16; P<.001; d=−0.74) and follow-up (mean 2.50, SD 1.32; t₄₃=−7.53; P<.001; d=−1.14), with 65% (31/48) of participants rated as at least minimally improved at postintervention and 75% (33/44) at follow-up (20/48, 42% and 25/44, 57%, respectively, rated much or very much improved). There was no significant difference between arms at postintervention (F_1,46=0.17; P=.68; η²_p=0.004) or follow-up (F_1,42=0.46; P=.50; η²_p=0.01).

Functional capacity, measured using the UPSA scale, showed a significant improvement over time (F_1,38.9=33.16; P<.001; η²_p=0.46, 90% CI 0.26-0.59). However, there were no significant effects for group (F_1,49.2=0.34; P=.56; η²_p=0.007), nor a group×time interaction (F_1,39.6=2.56; P=.12; η²_p=0.06). Post hoc analyses revealed significant improvement from baseline to postintervention in UPSA for both the active (14.19 points; 95% CI 8.60-19.79; t_38.8=5.13; P<.001) and the control (8.04 points; 95% CI 2.58-13.51; t_38.5=2.98; P=.01) groups.

Objective (diversity and frequency) and subjective (enjoyment and satisfaction) aspects of participation did not differ by time, group, nor group×time interaction. The full analyses of the ASAP can be found in Multimedia Appendix 7.

Composite scores (normed to the V1 data), comprising the individual scores of the 5 tasks (TNT, Stroop, Flanker, TaskSwitch, and SAAT), were computed for the ACE battery assessments, which were administered at V₁ and V₄. Composite scores significantly improved from baseline to postintervention (F_1,44.6=4.36; P=.04; η²_p=0.09). However, there were no significant effects of group (F_1,51.7=0.79; P=.38; η²_p=0.015), nor a significant group×time interaction (F_1,45.5=0.49; P=.49; η²_p=0.01). Post hoc analyses revealed that the changes between V1 and V4 were not significant for the active (−0.28 points; 95% CI −0.56 to 0.003; t_44.5=−1.99; P=.11) nor for the control arm (−0.14 points; 95% CI −0.42 to 0.15; t_44.5=−0.97; P=.67; Figure 3C).

Credibility did not differ between arms at either time point (week 1: LoH: mean 12.96, SD 7.22; LoH_control: mean 15.00, SD 6.38; t_47.7=−1.07; P=.29, 95% CI for the between-arm difference −5.88 to 1.80; V₄: LoH: mean 13.29, SD 7.47; LoH_control: mean 13.46, SD 5.95; t_44.0=−0.09; P=.93; 95% CI for the between-arm difference −4.04 to 3.70) and was stable within each arm over time, indicating that the matched control design sustained equivalent perceived credibility across arms. Credibility and outcome expectancy are distinct constructs, however, and comparable credibility did not entail comparable expectancy: outcome expectancy was higher in LoH_control than in LoH at week 1 (mean 12.87, SD 6.14 vs mean 9.29, SD 5.63; t_48.9=−2.17; P=.04; d=−0.61, 95% CI for the between-arm difference −6.89 to −0.27), indicating that control participants began the trial expecting more benefit; however, it did not differ between arms at V₄ (mean 10.75, SD 6.18 vs mean 9.35, SD 6.66; t_46.9=−0.77; P=.45; 95% CI for the between-arm difference −5.06 to 2.27). Both arms showed small numerical reductions in expectancy from week 1 to V₄, more pronounced in LoH_control (Δ=−1.99, paired t₂₃=−1.82; P=.08).

We exploratorily examined several factors as predictors of change in MADRS symptoms. First, the relations between changes in CC (using standardized composite scores) and clinical measures were assessed using residualized change scores, calculated by regressing postintervention scores on baseline scores for all measures. Changes in CC were not significantly associated with changes in depressive symptoms (MADRS: r=0.05, 95% CI −0.25 to 0.34; t₄₃=0.33; P=.74; PHQ-9: r=0.04, 95% CI −0.25 to 0.33; t₄₃=0.28; P=.78), quality of life (WHO-5: r=−0.12, 95% CI −0.40 to 0.18; t₄₃=−0.79; P=.44), or rumination (RRS-22: r=0.15, 95% CI −0.15 to 0.43; t₄₃=1.02; P=.31).

Next, we examined whether perceived choice (subscale of IMI) was associated with change in MADRS scores following treatment (see full results in Multimedia Appendix 10). In short, higher perceived choice significantly predicted greater MADRS improvement in the active arm only (B=1.41; adjusted P=.006), with this relationship reversed for the control arm.

Next, we examined whether blinding guesses were related to improvement in MADRS scores across the study population. The results showed that there was no significant difference in MADRS change between the active and control arms, regardless of participant guesses in the blinding check (see full analysis in Multimedia Appendix 10). Finally, the association between game progression metrics and change in MADRS and CC scores are summarized in Multimedia Appendix 11.