Leveraging Eye-Tracking Signals for Neurodegenerative Disease Classification with Deep Learning Models

Neurodegenerative diseases (NDs) can induce subtle changes in eye movements, which can be used for diagnosis or evaluation. This study employs deep learning models to classify NDs using raw eye-tracking data from 133 participants: 51 healthy controls (CTL), 25 with Alzheimer's Disease (AD), 39 with Parkinson's Disease (PD), and 18 with Parkinson's Disease Mimics (PDM)-PD-like diseases often misdiagnosed as PD. Eye movements were recorded during smooth pursuit, text reading, and picture description tasks. Results suggest the text reading task, being structured and cognitively demanding, best distinguished CTL from PD and AD, while the motor-oriented smooth pursuit task differentiated PD from PDM. Ablation studies suggest pupil size improved classification accuracy, especially for CTL vs. PD, and binocular data was crucial for distinguishing PD from PDM. This study shows deep learning models can classify NDs using raw eye-tracking data, becoming a potential tool for objective evaluation in neurology and primary care.Clinical relevance-This project illustrates the efficacy of deep learning methods in leveraging patients' eye movement to accurately predict the presence of neurodegenerative diseases.