All procedures involving human participants were conducted in accordance with the ethical standards of the Institutional Research Committee and with the Declaration of Helsinki. The study protocol was reviewed and approved by the Institutional Review Board (IRB) of the Korea Brain Research Institute (approval number: KBRI-202103-HR-02). Written informed consent was obtained from the legal guardians of all child participants prior to enrollment, and verbal assent was obtained from the children using developmentally appropriate language. Participants and their families were informed of their right to withdraw from the study at any time without penalty. Guardians received KRW 30,000 (~US $22) as compensation for participation and transportation expenses, and child participants received a small gift valued at ~KRW 20,000 (~US $15). All collected data were deidentified and stored securely to ensure confidentiality and privacy, and no identifiable images of participants were collected.

Paired-sample t tests indicated significant condition differences for all three behavioral measures. Accuracy was significantly higher in the color condition (t₁₂₆=3.81, d=0.34, 95% CI 0.03-0.08; P<.001). Reaction time was significantly shorter in the color condition (t₁₂₆=–7.72, d=–0.69, 95% CI −107.2 to −63.5; P<.001). Composite performance was significantly better in the color condition (t₁₂₆=6.81, d=0.62, 95% CI 0.003-0.006; P<.001). Effect sizes ranged from medium to large, suggesting that the interaction format had a more pronounced impact on manifest behavior than on observable neural network changes within this sample size. In contrast, neural indices showed no reliable differences between conditions. The task-related FC did not differ significantly (t₁₂₆=1.65, d=0.15, 95% CI −0.01 to 0.10; P=.10). Similarly, ∆FC (t₁₂₆=1.65, d=0.15, 95% CI −0.01 to 0.10; P=.10) and prefrontal global efficiency (t₁₂₆=1.52, d=0.14, 95% CI −0.01 to 0.05; P=.13) showed no significant differences, although all trended in the direction of higher values in the color condition.

Correlational analyses with BASC-2 Attention Problems T scores, a validated behavioral measure of attentional difficulties, showed that higher scores (indicating more attention problems) were associated with lower composite performance in the color condition (r=−0.20, 95% CI −0.37 to −0.03; P=.03), whereas the corresponding performance in the text condition was not significantly related (r=−0.05, 95% CI −0.23 to 0.12; P=.56). Fisher r-to-z transformation indicated that this difference approached but did not reach conventional significance (z=1.54, P=.12). Neural indices showed weak and nonsignificant correlations with attention problems in both conditions: color ∆FC (r=−0.07, 95% CI −0.24 to 0.10; P=.42), color global efficiency (r=−0.05, 95% CI −0.23 to 0.12; P=.55), text ∆FC (r=–0.14, 95% CI −0.31 to 0.03; P=.10), and text global efficiency (r=0.03, 95% CI −0.14 to 0.20; P=.73). Complete correlation coefficients are provided in Table 1.

To examine which features contributed to individual differences in performance within each interaction format, we estimated contributions using condition-specific random forest models with permutation-based variable importance (reported as normalized percentages; see the Methods section for details). Variables with higher importance scores contributed more to reducing the prediction error across decision trees.

In the color condition, age showed the largest contribution (58.86%), followed by cognitive measures including the PSI (17.06%) and the FSIQ (10.25%). Smaller contributions came from BASC-2 Attention Problems (4.79%), gender (4.75%), and neural indices (∆FC: 2.82%; global efficiency: 1.47%). The model achieved a cross-validated R² of 0.091, corresponding to ΔR²=0.100 relative to a mean-only baseline. Full importance values are provided in Table 2.

In the text condition, age again dominated (73.21%), with contributions from the PSI (8.53%), neural indices (∆FC: 6.34%; global efficiency: 4.40%), gender (3.51%), and the FSIQ (2.78%). BASC-2 Attention Problems (0.77%), and the WMI (0.47%) contributed minimally. The model achieved a cross-validated R² of 0.233, corresponding to ∆R²=0.274 relative to baseline.

Because age and the FSIQ contributed strongly in the original models, we conducted supplementary analyses using residualized outcome measures. Age and the FSIQ were first regressed out of composite performance (color: R²=0.271; text: R²=0.413), and the resulting residuals were used as dependent variables in new random forest models with neural indices and BASC-2 Attention Problems as predictors.

In the residualized color condition model, neural indices emerged as the largest contributors: global efficiency (39.79%) and ∆FC (35.75%), followed by BASC-2 Attention Problems (24.46%). In the residualized text condition model, a similar pattern emerged: global efficiency (43.60%), ∆FC (33.02%), and BASC-2 Attention Problems (23.37%). Residualized models achieved cross-validated R² values of 0.113 (color) and 0.202 (text), indicating that after controlling for age and general cognitive ability, neural and behavioral attention measures accounted for 11% to 20% of the remaining variance in performance. Complete residualized importance values are provided in Table 3.

This study examined how the interaction format influences cognitive load and measurement validity in a tablet-based Stroop task for children. Using within-subject comparisons between color-based and text-based response formats, combined with prefrontal fNIRS and clinical behavioral measures, we addressed two primary questions: whether the interaction format affects cognitive load and performance and whether performance differences translate into differential clinical validity.

Behavioral results clearly demonstrated the effects of interaction formats. The color-based response condition yielded significantly faster responses, higher accuracy, and better composite performance compared to the text-based response condition (all P<.001, with medium-to-large effect sizes). Critically, color-based performance showed a significant correlation with parent-reported attention problems (BASC-2; r=−0.20, P=.03), whereas text-based performance did not (r=−0.05, P=.56). Although the relatively wide CIs suggest that the precise magnitude of these associations warrants replication, this differential pattern indicates that the color format showed a pattern more consistent with clinical validity by better isolating attentional control from other cognitive demands. Random forest analyses further revealed that after controlling for age and general cognitive ability, neural indices of prefrontal efficiency (global efficiency, ∆FC) emerged as the dominant predictors of individual performance differences in both formats, accounting for ~35% to 44% of residual variance.

The observed performance differences are consistent with the intended manipulation of cognitive load. The text-based response format required dual semantic processing: participants had to ignore the word meaning while simultaneously mapping the ink color to a written label [33]. In contrast, the color-based response format preserved the core Stroop conflict while eliminating competing verbal response options, thereby reducing extraneous cognitive load [45]. Consistent with this interpretation, the color condition showed not only superior behavioral performance but also descriptively higher prefrontal efficiency (∆FC, global efficiency). Although these neural differences did not reach statistical significance, such small-to-moderate effects are consistent with the typical magnitude observed in fNIRS research with developmental populations [65]. In addition, PFC maturation continues throughout childhood, and the developing PFC networks in our sample may not yet exhibit sufficiently stable connectivity patterns to produce detectable condition-level differences [66]. Furthermore, because both conditions shared the core Stroop conflict and differed only in the response format, the neural workload differences may have been too subtle to detect at the group level.

It is worth noting that modifying the response format may have also altered the relative contribution of different cognitive processes at the response selection stage. However, the preservation of the core Stroop conflict at the stimulus level and the significant association between color-based performance and clinical attention measures suggest that the fundamental inhibitory control construct was maintained across both formats. Moreover, color-based formats may be especially appropriate for children in early literacy stages, for whom text-based responses could introduce construct-irrelevant variance related to reading ability rather than attentional control [35].

As an exploratory analysis, random forest analyses provided a complementary view of how individual differences relate to task performance in each format. In the strict models, age had the largest contribution in both formats, and cognitive measures, such as the PSI and FSIQ, accounted for much of the remaining signal, with neural features (prefrontal global efficiency, ∆FC) contributing smaller percentages (Table 2). Because age and IQ dominated these models, we conducted residualized analyses that removed their variance from the outcome prior to modeling. In these residualized models, neural indices carried most of the remaining predictive signal in both formats, with BASC-2 Attention Problems accounting for the remainder (Table 3).

These findings highlight two key points. First, the higher contribution of BASC-2 Attention Problems in the color condition is consistent with the stronger correlation observed between color-based performance and the clinical attention benchmark, suggesting that the lower-load interface better isolates attentional control from other developmental influences. Second, the dominant role of age and IQ in the strict models underscores the need to treat these variables as covariates in attentional research, echoing prior findings that developmental and general cognitive factors exert a substantial influence on performance [35,67]. Notably, however, the relative contribution of age was smaller in the color format (58.86%) than in the text format (73.21%). This pattern suggests that reducing semantic interference was associated with lower cognitive load indicators and a smaller relative contribution of developmental differences dominating performance, thereby providing a clearer assessment of attentional control.

At the same time, format-specific nuances emerged that distinguished average performance from individual variability. In terms of mean values, both ∆FC and global efficiency were descriptively higher in the color condition, which is consistent with more efficient prefrontal organization when cognitive demands are lower [61]. However, random forest analyses revealed a complementary pattern: in the text condition, individual differences in these same neural indices showed higher predictive importance, indicating that variability in neural efficiency became a stronger determinant of performance under elevated cognitive load.

This pattern reflects two distinct aspects of neural function. First, the color format was associated with descriptively higher average neural efficiency across all children, consistent with reduced cognitive demands. Second, the text format was associated with a greater role of individual neural capacity. When semantic interference increased, the children’s ability to maintain efficient prefrontal processing became more critical for successful performance. This aligns with cognitive load theory, specifically the reduction in extraneous load, while preserving intrinsic task demands, which posits that when extraneous demands are high, individual capacity for managing cognitive resources becomes a limiting factor [24]. Prior evidence supports this interpretation, showing that elevated cognitive load disrupts inhibitory processing and exacerbates individual differences in performance [68].

The relatively modest cross-validated R² values (modest cross-validated performance in residualized models) reflect the inherent complexity of predicting individual performance from neural and psychometric measures alone. These models were designed to estimate the relative contribution of heterogeneous features rather than to achieve precise behavioral forecasting [69]. From this perspective, the key insight is not the absolute predictive accuracy but rather the shift in variable importance: neural indices became dominant predictors once age and general cognitive ability were controlled, accounting for 35% to 44% of residual variance. This underscores that PFC network efficiency captures performance variance distinct from broader developmental and intellectual factors.

Taken together, these findings have direct implications for the design of digital assessment tools for children. For clinical diagnostic guidelines, age-based normative standards should serve as a primary framework, given the dominant role of age in predicting performance. However, age-based norms and interface optimization are complementary: although norms ensure appropriate score interpretation across developmental stages, optimized response formats can reduce construct-irrelevant variance and improve measurement precision at every age level. The observation that the interaction format influences behavioral performance and clinical validity, with suggestive trends in neural efficiency, underscores the need to explicitly consider cognitive load when designing Stroop-based assessments. Specifically, reducing extraneous semantic demands through color-based response formats may enhance both measurement accuracy and ecological validity in pediatric populations.

This study has several limitations. First, although the within-subject design controlled for stable individual differences, the fixed task order (color followed by text) introduces the possibility of practice or fatigue effects. However, practice effects would be expected to benefit the second condition, potentially reducing rather than inflating the observed differences. The fact that the color condition showed superior performance despite being administered first suggests that the format effects were robust to any potential order confounds. Second, the sample comprised typically developing children aged 6-12 years, limiting the generalizability to broader developmental stages and clinical populations. Third, the neural feature set was necessarily coarse, relying on global efficiency and condition-level ∆FC; finer-grained metrics, such as event-locked or channel-specific analyses, may offer greater sensitivity. Finally, inferences about feature contributions relied on random forest models with modest predictive performance (R²=9%-23%). These results should be regarded as exploratory estimates of relative importance rather than robust predictive models. Nonetheless, they provide a valuable starting point for understanding the complex interplay between interface design and neurocognitive function in children.

Future work should (1) counterbalance the task order and include repeated sessions to assess reliability; (2) extend the age range and incorporate clinical subgroups to evaluate external validity; (3) use more sensitive neural measures, including time-resolved connectivity and channel-specific analyses; and (4) apply mediation or structural equation models in preregistered analyses to formally test whether neural indices statistically account for format-related performance differences.

The findings of this study demonstrate that the interaction format influences both cognitive load and measurement validity in tablet-based Stroop assessments for children. Color-based response options, which reduce extraneous semantic processing demands, yield better behavioral performance and stronger correlations with clinical attention measures compared to text-based options. Random forest analyses further revealed that neural indices of prefrontal efficiency become the dominant predictors of individual differences once age and general cognitive ability are controlled.

These results underscore the importance of interaction format calibration in pediatric digital assessments. To the best of our knowledge, this study provides some of the first empirical evidence directly comparing how the response format affects both neural indices of cognitive load and clinical validity in a pediatric sample, extending cognitive load theory from educational settings into digital health assessment design. As digital cognitive assessments become more prevalent in research and clinical settings, careful attention should be paid to how interface design shapes cognitive load for ensuring that these tools are both developmentally appropriate and clinically valid. For Stroop-based attention assessments in children, this work offers initial guidance: prioritizing response formats that minimize extraneous cognitive demands may enhance both measurement fidelity and clinical utility.