Multimodal Temporal Deep Learning for ADHD Classification in Ecologically Valid Virtual Reality: An Exploratory

Abstract

Attention Deficit Hyperactivity Disorder (ADHD) assessment increasingly incorporates artificial intelligence (AI), yet most approaches rely on laboratory-derived or questionnaire-based data with limited ecological validity. Virtual reality (VR) enables controlled simulation of complex, goal-directed activities, but existing VR-based tools largely exploit aggregated task-performance metrics and underutilize the rich multimodal signals available during interaction.

This exploratory study investigates how representation design and model architecture influence ADHD classification within an ecologically grounded Multiple Errands Test implemented in immersive VR. We address three research questions: (RQ1) Do multimodal behavioral and physiological signals provide additional discriminative value beyond traditional task metrics? (RQ2) Does preserving temporal structure improve classification performance compared to global aggregation? (RQ3) Do temporal deep learning architectures outperform classical machine learning models when applied to multimodal VR-derived sequences?

Twelve adults (5 ADHD, 7 controls) completed the task. We compared global aggregated features, window-based aggregated representations, and raw multichannel temporal sequences using Logistic Regression, Support Vector Machines, Random Forest, Gradient Boosting, and Temporal Convolutional Networks (TCNs), under Leave-One-Subject-Out cross-validation.

Results show that multimodal signals provide limited benefit under global aggregation but yield improved performance when temporal dynamics are preserved. TCN-based models achieved the highest accuracy (up to 0.79), suggesting that sequence-aware architectures better exploit irregular, action-aligned dynamics. These findings highlight the critical interaction between representation design and model architecture in multimodal time-series classification within ecologically valid VR assessments.