Generative Artificial Intelligence-Enabled Saliency Analysis of Eye Tracking in Cerebral/Cortical Visual Impairment...

Children with cerebral/cortical visual impairment (CVI) have neurological conditions that impact visual pathways in the brain, leading to deficits in both lower- and higher-order visual function that can be challenging to measure, especially in children with neurodevelopmental delays. We developed a method that combines eye tracking with generative artificial intelligence (AI) to characterize attention to lower- and higher-level visual characteristics in children with CVI.

A cross-sectional, comparative cohort study.

Forty-three children with CVI and 40 typically developing pediatric controls.

Participants viewed a series of still images while gaze location was recorded by an eye tracker. For each image, saliency maps highlighting 15 lower- and higher-level visual features were created by a combination of generative AI and pixel-based techniques. Participants' eye tracking fixation maps were compared to saliency maps to calculate fixation saliency values. Fixation saliency values were then compared between CVI participants and controls. The 95% prediction interval of control fixation saliency values was used to establish a normative range, and each CVI participant's fixation saliency values were categorized as normal or abnormal. The distribution of normal and abnormal fixation saliency values for each participant was termed the saliency analysis of eye tracking in CVI (SET-CVI) signature.

Fixation saliency values in children with CVI and controls; accuracy, sensitivity, specificity; and area under the receiver operating characteristic curve (AUC) of SET-CVI in classifying children with CVI and controls.

Children with CVI had lower fixation saliency values for higher-level visual features such as human faces and bodies (P ≤ 0.0065). Children with CVI had higher fixation saliency values for lower-level visual features such as certain colors (P ≤ 0.04). Saliency analysis of eye tracking in CVI had 98% accuracy, 100% sensitivity, and 95% specificity in classifying children with CVI and controls, with an AUC of 0.98 (P < 0.0001).

Eye tracking combined with generative AI can quantify attention to lower- and higher-level visual characteristics in children with CVI and distinguish affected children from controls. Saliency analysis of eye tracking in CVI signatures may be useful for guiding individualized interventions, tracking progress longitudinally, and serving as an outcome measure in future clinical trials to identify an evidence-based medical treatment for CVI.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.