Real-world clinical validation of brainstem-based ocular biomarkers for ADHD classification in children and adults.
TL;DR
This work proposes an objective deep-learning approach classifying ADHD via task-evoked pupil diameter and binocular eye-movement synchrony during a visual cueing task in 439 participants across 14 clinical centers, offering a robust tool to reduce subjective clinical bias.
Abstract
Current ADHD diagnostic practices rely on subjective rating scales and continuous performance tests with limited specificity. We propose an objective deep-learning approach classifying ADHD via task-evoked pupil diameter and binocular eye-movement synchrony during a visual cueing task in 439 participants across 14 clinical centers. We implemented two independent models: a multiple instance learning (MIL) framework for pupil dynamics and conventional classifiers for eye-movement synchrony. The outputs of these models were fused to derive two novel indices, a diagnostic score and an impulsivity score. Using a three-zone policy (healthy, ADHD, uncertain) to manage diagnostic uncertainty, pediatric cross-validation (N=324) yielded diagnostic and impulsivity sensitivities of 0.79 and 0.74, and specificities of 0.82 and 0.70. Adult external testing (N=115) achieved specificities of 0.86 and 0.92, with sensitivities of 0.66 and 0.68. Explainable AI confirmed predictions are driven by increasing pupil responses immediately following cue and stimulus onsets. Statistical projections estimate that integrating this tool with standard rating scales can optimize diagnostic pathways, yielding 95% sensitivity in a screening mode or 96% specificity in a confirmation mode. These physiologically grounded biomarkers reliably quantify cognitive impairments, offering a robust tool to reduce subjective clinical bias.