A total of 32 children aged 10-14 years participated in this study. This age range was selected to ensure participants had the developmental capacity to understand instructions, operate the VR equipment, and adhere to the training protocol. A total of 22 participants were recruited from the Department of Psychiatry at a general hospital in Seoul, where they were diagnosed with ADHD by a child psychiatrist. None had comorbidities. Notably, 18 participants were medicated with varying types and dosages, while 4 were medication-naive. An additional 10 participants were recruited from the community via social networks and community boards. Inclusion criteria for community participants were based on the mobile-based cognitive control assessment app, CoCon. To be included, candidates had to score below a T-score of 40 on at least one of 4 indices: sustained attention, working memory, cognitive control, and cognitive execution. A T-score of 40 on the CoCon app represents a mean score 50 (1 SD) below the mean. There were no significant differences between the clinical and community participants in the basic variables (Table S1 in Multimedia Appendix 1).

Sample size calculation was performed using G*Power version 3.1 software (Heinrich-Heine-Universität Düsseldorf [31]). Initially, the study was designed as a between-subjects repeated measures ANOVA, and a sample size of 86 was recommended, assuming a medium effect size (Cohen f=0.25), an α level of .05, and a statistical power of 0.80. However, due to uncontrollable circumstances, such as the COVID-19 pandemic, it was not feasible to recruit enough participants for the between-subjects design. Therefore, the study design was changed to a within-subjects repeated measures ANOVA. Based on this revised design, a sample size of 28 was recommended under the same parameters (Cohen f=0.25; α=.05; power=0.80), assuming a medium effect size for a within-subjects context. To account for an estimated dropout rate of 20%, the final target sample size was increased to 32 participants. Overall, 3 participants dropped out during training, resulting in a final count of 29 participants. In addition, 1 participant who reported red-green color blindness after the pretraining test had their Stroop test data replaced with mean values from corresponding variables for analysis. For the 29 participants who completed all study assessments, all of them also completed the full 20-day training protocol, demonstrating 100% adherence among this group.

This VR game was developed using the Unity (Unity Technologies) game engine and is compatible with the Oculus Quest 2 (Meta) VR device. The VR game app was designed for home-based cognitive control training, with each participant expected to engage in at least 20 minutes of training daily for 20 days. Research assistants monitored training data and proactively addressed difficulties over the phone each day, conducting in-home visits as needed for technical challenges.

This study included 3 assessment phases comprising a pretest for initial assessment, a posttest after 20 days of training, and a follow-up test conducted 3 months later. To evaluate the retest effect specifically, 14 participants underwent an additional test (retest) without training. Participants, including those in the no training retest group, were alternately allocated based on enrollment order with randomization. After evaluations, this group participated in training sessions, followed by subsequent posttests and follow-up assessments.

The pretest included the Korean Wechsler Intelligence Scale for Children-Fourth Edition (K-WISC-IV) [32], the Korean Color Trails test (CTT) 1 and 2 [33], the Korean Stroop test [34], the NIH toolbox [35], and the Korean Child Behavioral Checklist (K-CBCL) [36], lasting approximately 90 minutes excluding the parent-completed K-CBCL. The posttest and follow-up tests included CTT 1 and 2 [33], Stroop test, Flanker test from the NIH toolbox [35], and the K-CBCL [36], each taking about 30 minutes, excluding the parent-completed K-CBCL [36]. Training monitoring and assessments were conducted by research assistants who had graduated from master’s programs specializing in clinical psychology.

Data analysis was performed using Jamovi software [38]. Demographic data were summarized by calculating means, SDs, and frequencies. Game-related behaviors were analyzed, including login frequency, instances of exiting games midway, total game time, total trial numbers, correct-incorrect counts, and reaction times. Paired 2-tailed t tests evaluated retest effects. The main analysis used repeated-measures ANOVAs to assess training effects at 3 time points, such as pretraining, posttraining, and 3 months posttraining. Scatter plots depicting pretest, posttest, and follow-up test results were generated using the ggplot function [39,40] in Jamovi software [38]. In addition, the Pearson correlation analysis was conducted between game-related behaviors and difference scores between pretests and follow-up tests in the Stroop test and CBCL total score, attentional problem score, and ADHD score. Training response patterns were analyzed using the k-means clustering method, applying Lloyd algorithm to identify distinct clusters based on standardized z scores (mean 0, SD 1) from difference scores between pretests and follow-up tests in the Stroop test and CBCL total score, attentional problem score, and ADHD score. Group comparisons were conducted using independent 2-tailed t tests across various parameters, including IQ subscales (verbal comprehension, perceptual reasoning, working memory, and processing speed) and game behaviors (correct/incorrect ratios, total response times, total trial numbers, number of exits during games, and login frequencies).

This study received approval from the institutional review board (SWU IRB-2020A-56; Multimedia Appendix 2). Written informed consent was obtained from participants and their parents, with each participant receiving 90,000 KRW (US $70) upon completing all evaluations. Data were anonymized.

Basic data analysis indicated that the mean age of the participants was 11.3 (SD 1.13) years, with ages ranging from 10 to 14 years (no retest group: 11.1, SD 1.28 years; no training retest group: mean 11.5, SD 0.94 years). A visual representation of the entire process depicting subsequent posttests and follow-up assessments can be found in Figure 2. The mean total IQ (K-WISC-IV) was 94.2 (SD 16.5; no retest group: mean 95.4, SD 15.97; no training retest group: 93, SD 17.62). The gender distribution among the 29 particpants comprised 21 (72%) boys and 8 (28%) girls. There was no significant group difference regarding gender ratio, with a chi-square (χ²) test resulting in P=.07.

The retest effects, assessed using paired t tests, revealed no significant changes between initial assessments and reevaluations without intervention across various measures (t₁₃=−1.979, P=.07 for CTT 1; t₁₃=−0.711, P=.49 for CTT 2; t₁₃=−0.899, P=.39 for Flanker test from NIH toolbox; t₁₃=−1.773, P=.10 for Stroop test; t₁₃=1.224, P=.30 for CBCL Total score; t₁₃=1.694, P=.11 for CBCL Attention Problems score; and t₁₃=1.224, P=.24 for CBCL ADHD score).

Table 1 details the observed training effects across pretest, posttest, and follow-up assessments. Repeated measures ANOVA revealed significant effects in the Stroop test (F_2,56=4.97; P=.01; ηp²=0.151; pretest<posttest P=.006), CBCL Total Problems Score (F_2,56=21.0; P<.001; ηp²=0.429; pretest<posttest P<.001; pretest<follow-up test P<.001), CBCL Attention Problems Score (F_2,56=11.7; P<.001; ηp²=0.294; pretest<posttest P<.001; pretest<follow-up test P=.002), and CBCL ADHD score (F_2,56=3.46; P=.04; ηp²=0.110; pretest<posttest P=.03). Figure 3 presents plots comparing the within-group results from the pretest, posttest, and follow-up tests, created using ggplot2. CTT1 and CTT2 and the Flanker test from the NIH toolbox did not demonstrate any significant effects among the pretest, posttest, and follow-up test.

A Pearson correlation analysis revealed significant positive correlations between the correct response ratio and the difference scores on the CBCL Total Problems, Attention Problems, and ADHD scales. Furthermore, significant correlations were found between the mean correct response time and the difference scores on the Stroop Color-Word test, as well as the CBCL Attention Problems and ADHD scales. The total login frequency was also significantly correlated with the difference scores on all 3 CBCL scales (see Table S1 in Multimedia Appendix 3 for details). The k-means cluster analysis, which used z scores of the difference between pretest and follow-up on the Stroop test and CBCL scales, produced 2 distinct clusters. Cluster 1 (n=10) did not show significant improvement, while cluster 2 (n=19) exhibited overall improvements from pretest to follow-up. A comparative analysis did not reveal any significant differences between the 2 clusters in age or on the IQ-related indices. However, there was a borderline significant difference in gender distribution between the clusters (χ²₁=3.84; P=.05). Furthermore, a significant association was found between sample type (clinical vs community) and cluster membership (χ²₁=8.03; P=.005) (see Tables S1 and S2 in Multimedia Appendix 4 for details). In addition, there were significant differences in game-related behavioral variables, including mean correct response time (t₂₇=−2.56; P=.02) and the correct response ratio (t₂₇=2.60; P=.02). These results suggest that a longer correct response time and a lower correct response ratio were associated with better training outcomes. The detailed results are presented in Table 2 and Figure 4.

This pilot study provides preliminary evidence that a novel, at-home VR cognitive training game can significantly improve cognitive control and parent-reported ADHD symptoms in children. Notably, these therapeutic gains were sustained at a 3-month follow-up, and an analysis of a nonintervention group confirmed that the improvements were not solely attributable to test-retest effects. The study also demonstrated high feasibility. The integration of a telemonitoring system was effective in maintaining excellent adherence to the 20-day protocol, overcoming common challenges associated with remote interventions.

However, the Flanker test from the NIH toolbox and the CTT did not show significant training effects. This lack of significant findings may stem from the nature of these tasks, which primarily assess more automatic attentional processes, in contrast to the Stroop test, which targets controlled attention. Prior research has suggested that basic attention networks can exhibit unstable changes with repeated testing, making such measures potentially less sensitive to short-term training improvements [41]. A k-means clustering analysis was conducted based on changes observed in the Stroop Color-Word test, CBCL total behavioral scores, attentional problems, and ADHD symptom scores. This analysis revealed distinct patterns of training response among the identified clusters. Overall, the 2 clusters displayed contrasting trends: one group demonstrated a positive training effect, while the other showed no clear improvement. Notably, there were no significant differences in age and IQ indices between the clusters, suggesting that age and general intelligence did not account for the observed variability in training outcomes. Interestingly, the positively responding cluster exhibited longer correct response times and lower correct response ratios compared to the nonresponsive group. Although these findings appear counterintuitive, they may reflect greater engagement or more sincere participation, indicated by a higher number of total trials and longer reaction times, which could underlie the observed improvements. Notably, a significant association was found between the participant source (clinical vs community) and training response, with the community sample being less responsive to the training than the clinical sample. While the baseline assessment showed no significant differences between the 2 groups in IQ, age, or attention-related problems, the difference in training outcomes may suggest lower motivation in the community sample. However, given the small size of the community group, this interpretation is speculative and requires further investigation.

Despite these compelling findings, several critical limitations must be acknowledged. First, the study lacked a control group, which limits the ability to draw causal conclusions about the effects of the training. The within-subjects design, without a parallel control, restricts the interpretation of training efficacy in a more systematic and rigorous manner. Second, the relatively small sample size (N=29), along with imbalanced participant characteristics, particularly in terms of gender distribution and medication status, reduces the generalizability and statistical power of the findings. Third, the study did not assess higher-order cognitive functions, such as executive functioning. Incorporating such measures could have provided a more comprehensive understanding of the relationship between neuropsychological test performance and parent-reported behavioral outcomes.

In conclusion, the newly developed VR-based cognitive control game, enhanced with a modified adaptive staircase algorithm and supported by a telemonitoring system, demonstrated its potential as an accessible and effective at-home intervention for children with ADHD symptoms. Building on research showing that user characteristics influence engagement with tele-mental health services [42], an important next step is to explore which specific user profiles benefit most from this VR-based intervention.