Machine learning-driven analysis of temporal pupil dynamics for interpretable ADHD diagnosis

Attention-deficit/hyperactivity disorder (ADHD) is a prevalent neurodevelopmental disorder characterized by inattention and hyperactivity. Current diagnostic methods rely on bias-prone subjective assessments, such as clinical interviews and behavior rating scales. Objective biomarkers remain elusive hindering standardized ADHD diagnosis. Pupillometry, measuring pupil responses linked to cognition and attention, offers a promising, objective alternative. However, prior work often overlooks clinically relevant features and lacks interpretability, limiting clinical adoption. We introduce an interpretable machine-learning framework leveraging temporal pupil dynamics to classify ADHD and control groups. The primary novelty of our work lies in identifying and statistically validating task-aligned features-specifically, novel dynamic pupil dilation and constriction rates extracted in block-wise temporal segments-which capture subtle attentional fluctuations overlooked by prior models. We analyzed published pupillometry data from 49 participants (21 controls, 28 ADHD, 17 assessed on and off medication) during a visuospatial working memory task. Candidate features were identified through statistical analyses using mixed analysis of variance. Classification models were trained to prioritize interpretability by utilizing statistically significant, literature-supported features. Model transparency was enhanced with heatmaps and feature-importance charts. The models demonstrated strong classification performance: using pupil features alone yielded 84.4% accuracy (area under the receiver operating characteristic (AUROC) 88.6%). Including task performance improved accuracy to 86.7% (AUROC 91.5%). Final integration of reaction time metrics achieved 88.9% accuracy (AUROC 90.8%), with 97.8% sensitivity and 82.2% specificity. By leveraging interpretable, dynamic pupil metrics, our approach advances objective, reproducible ADHD diagnosis and supports clinical deployment.