A Novel Approach Using Serious Game Data to Predict the WISC-V Processing Speed Index in Children With Attention-Deficit/Hyperactivity Disorder: Machine Learning Study.

Background: The processing speed index (PSI) of the Korean Wechsler Intelligence Scale for Children-Fifth Edition (K-WISC-V) is highly correlated with symptoms of attention-deficit/hyperactivity disorder (ADHD) and is an important indicator of cognitive function. However, restrictions on the frequency of testing prevent short-term PSI assessments. An accessible, objective technique for predicting PSI scores would enable better short-term monitoring and intervention for children with ADHD. Objective: To enable objective and accessible monitoring of cognitive function beyond traditional clinical assessments, this study aimed to develop a machine learning model that predicts the PSI scores of children with ADHD using behavioral data from serious games. Methods: Sixty-eight children (6-13 y of age) with ADHD were recruited, and after excluding incomplete data, 59 participants were included in the final analysis. The participants completed an initial PSI assessment using the K-WISC-V followed by 25 minutes of engagement with serious game content. Data from the game sessions were used to train machine learning models, and the models' performance in predicting PSI scores was evaluated using the root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percent error (MAPE), with K-fold cross-validation (k=4) applied to ensure robustness. Results: Among the individual machine learning models, support vector regression (SVR) had the best performance, with the lowest RMSE of 11.288, MAE of 7.874, and MAPE of 7.375%. The best overall performance was achieved by the ensemble integrating AdaBoost, Elastic Net, and SVR, which recorded the lowest RMSE of 10.072, MAE of 6.798, and MAPE of 6.611%. The predictive accuracy of this ensemble model was highest for PSI scores near the mean value of 100, demonstrating its reliability for clinical applications. Conclusions: The developed PSI prediction model has the potential to serve as an objective and accessible tool for monitoring cognitive function in children with ADHD. As a complement to traditional assessments, this approach allows continuous tracking of symptom changes and can support more personalized treatment planning in both clinical and everyday settings, which may improve accessibility and adherence. However, the findings need to be validated in larger, more diverse populations, and the long-term feasibility of using serious games in clinical and educational settings must be further examined.