Establishing Objective Ground-Truth for PediatricADHD Engagement: A Methodological Frameworkand Benchmark Dataset

Detecting engagement in pediatric ADHD typically relies on subjective, biasedhuman labeling. To address this, we introduce an objective annotation framework utilizing behavioral performance from the Integrated Visual and AuditoryContinuous Performance Test (IVA-2). From 21 children, we derived a continuousEngagement Score (Es) and windowed categorical labels, validating them againstofficial IVA-2 clinical scales. The proposed score demonstrated excellent internalreliability (r = 0.84), strong convergent validity (AUC = 0.97), and clinical sensitivity, significantly differentiating ADHD severity groups (Kruskal–Wallis, p <0.05). Temporal analysis revealed a strong concurrent correlation (r = 0.57) withreal-time dynamics. Benchmarking established that static Random Forest modelscurrently outperform temporal baselines, with eye-gaze dynamics serving as thestrongest unimodal predictor. We release the synchronized video features (OpenFace 2.0) and validated labels, providing a robust, privacy-preserving benchmarkfor objective engagement detection in clinical and educational settings.