Saliency Models Reveal Reduced Top-Down Attention in Attention-Deficit/Hyperactivity Disorder: A Naturalistic Eye-Tracking Study

Abstract

Objective: Attention-deficit/hyperactivity disorder (ADHD) is a highly prevalent neurodevelopmental disorder. It is associated with deficits in executive functions, especially invisual attention. Deviant visual attention in ADHD is suspected to arise from imbalancesbetween top-down and bottom-up mechanisms. However, it is unclear which of thesemechanisms propels the aberrant visual attention.Method: Here, we investigated differences in visual attention in ADHD and neurotypical controls(NC) using eye-tracking in a naturalistic video viewing task in 815 medication-naï ve children andadolescents (age range 5–21 years). We used two opposing saliency models: Finegrained, basedon low-level image features, was chosen to estimate bottom-up visually relevant areas. ViNet, ahigher-level saliency model based on deep neural networks and trained on the gaze of NCparticipants, was selected to determine top-down visually relevant regions. Correspondencebetween gaze and both saliency maps was calculated using normalized scanpath saliency (NSS),thus measuring the extent of coherence to bottom-up and top-down relevant contents.Results: Individuals with the combined presentation of ADHD (ADHD-C) showed lower meanNSS for the top-down saliency map, but not the bottom-up one, compared to NC. This contrastindicates poorer top-down control as a major contributor to impaired visual attention in ADHD-C.There was no significant effect for the ADHD predominantly inattentive presentation group.Conclusion: This study demonstrated the use of eye-tracking for differentiating between top-downand bottom-up visual attention. It shows that in ADHD-C, a reduction of top-down visual attentionis key to an impaired competition between bottom-up and top-down visual attention.