Leveraging Persistent Homology of Eye Movements for Neural Disorder Screening

Abstract

Abnormal ocular behaviors are associated with numerous neural disorders, detectable through specific eye movement patterns. Eye tracking--a non-invasive, accessible method--has thus been investigated as a tool for the automated diagnosis of these disorders. However, traditional eye movement feature extraction methods have limitations and are not universally applicable across diverse tasks. In this study, we present a novel feature extraction approach using Persistent Homology, a topological data analysis technique, to capture spatial and temporal topological features from eye movement data. Topological features retain multiple theoretical properties, making them robust for identifying abnormal eye movements and broadly applicable across tasks. By using only spatio-temporal topological features, our approach demonstrated promising performance in screening dyslexia and ADHD with a simple shallow perceptron classifier containing one or two hidden layers. These results support the informative validity of topological features as well as the efficacy of our proposed methods for extracting them.