Artificial Intelligence Based Ocular Motor Digital Biomarkers for Neurologic Disease Phenotyping

Neurological disorders impact a large percentage of the global population and are a vast area of research in the clinical field. However, state of the art diagnostic measures such as MRI and CT scans are invasive and expensive. Saccades, rapid fixations in eye movements, are a promising but underutilized non-invasive biomarker for neurological abnormalities due to limited publicly available data and privacy concerns. To address this, we developed a pose-guided video generation model that produces synthetic saccades of three types: normal, bilateral hypermetria, and bilateral hypometria that mimic real eye movement patterns observed in clinical settings. We trained an MViT-V2 video classification model on the synthetic data as a baseline and tested its performance on clinical saccade data. Our approach demonstrates the potential of synthetic data to enable accurate and scalable saccade-based diagnostics, reducing the dependency on invasive imaging.