Classifying ADHD Presentations Using Temporally Segmented Behavioral Data from a Serious Game

Distinguishing attention deficit/hyperactivity disorder (ADHD) presentations, such as predominantly inattentive (ADHD-I) and hyperactive/impulsive (ADHD-HI), is clinically important for management. However, traditional ADHD assessments often rely on subjective evaluations, potentially introducing biases and limiting diagnostic accuracy. As a more objective approach, this study aimed to develop a machine learning-based classification model using behavioral data derived from serious games performed by children diagnosed with ADHD. A total of 51 children aged 6 to 13 participated, and both parent-reported Korean-ADHD Rating Scale scores and serious game performance data were collected. Behavioral data collected from serious games were temporally segmented into early and late phases of gameplay, and discriminative behavioral features were identified through a machine learning-based feature selection process. Three machine learning models, Random Forest, Support Vector Machine, and eXtreme Gradient Boosting, were trained and evaluated. Models using temporally segmented behavioral features showed strong classification performance, particularly for the ensemble-based models. Among them, the Random Forest model demonstrated the best test performance, achieving an accuracy of 81.818%, an F1-score of 85.714%, and an AUROC of 83.333%. Furthermore, behavioral differences between the ADHD-I and ADHD-HI groups became more pronounced during the latter half of gameplay. In particular, children with ADHD-I showed longer decision times, longer times to the final touch, more unnecessary touches, and lower maximum difficulty levels than those with ADHD-HI, indicating that cognitive processing becomes slower and response strategies become less efficient as gameplay progresses into the later phase. These findings indicate that models using serious game-derived behavioral data, particularly with temporal feature engineering, can effectively classify ADHD presentations and may serve as a scalable, quantitative tool to support clinical assessment and the development of personalized interventions.