Analyzing Attention Patterns in Children with ADHD through Interaction Data from Educational Games

Abstract

Interaction data generated by serious games provide an opportunity to characterize cognitive behavior at a substantially finer temporal resolution than conventional session-level performance scores. However, game-based ADHD research has largely focused on aggregate outcomes or participant classification, while the sequential organization and within-task dynamics of attention-related behavior remain comparatively underexplored. This study investigates attention-related behavioral patterns derived from Attention Slackline and Attention Robots in the BALLADEER dataset. Among 85 children aged 6-12 years, the primary comparison cohort comprised 58 participants with diagnosed=yes and 18 with diagnosed=no; nine participants with diagnosed=undetermined were excluded from primary group comparisons. Across 254 Attention Slackline sessions, 4,820 of 4,822 expected target opportunities were successfully reconstructed. Mixed-effects models were used to analyze omission, commission, and reaction time (RT), while a Hidden Semi-Markov Model (HSMM) was used to infer latent behavioral states from sequential interaction data. Two machine-learning models were evaluated using nested five-fold participant-level cross-validation, and cross-game consistency was examined in 70 participants with sufficient data from both tasks.

The ADHD group showed higher odds of omission than the comparison group (OR=2.51; 95% interval: 2.25-2.80) and higher odds of commission (OR=1.83; 1.60-2.09), whereas no reliable ADHD-related difference in RT or ADHD-specific increase in omission over time was observed. The HSMM identified three latent behavioral states. An omission-dominant state showed higher adjusted occupancy in the ADHD group (+7.0 percentage points; p=0.0325), although the effect did not remain significant after false discovery rate correction (q=0.068). Out-of-fold AUROC ranged from 0.53 to 0.59, and temporal features did not improve discrimination relative to conventional behavioral features. Cross-game associations between Attention Slackline and Attention Robots were generally weak.

These findings indicate that game interaction data can reveal attention-related behavioral differences at both event and latent-state levels, but do not currently support accurate ADHD classification or strong generalization across tasks. The primary contribution therefore lies in behavioral computing and temporal behavioral modeling rather than automated ADHD diagnosis.