AI-based Classification of Parkinson’s Disease Using Quantitative Oculomotor Biomarkers
Ana Coito1, Erika Han2, Hassan Fadavi3, Bruno Hauser1, Pia Massatsch1, Valentina Stozitzky4, Dongli Li4, Bettina Balint2, Helen Dawes3, Konrad P. Weber5
1machineMD, 2Department of Neurology, University Hospital Zurich, 3University of Exeter, 4gaitQ, 5Departments of Neurology and Ophthalmology, University Hospital Zurich
Objective:
The development of artificial intelligence (AI) models for diagnosing Parkinson’s disease (PD) using ocular motor parameters obtained from a virtual reality (VR)-based medical device.
Background:
PD diagnosis currently relies largely on clinical examination, underscoring the need for objective biomarkers to enhance diagnostic accuracy, early detection and disease monitoring. Eye movement and pupil function abnormalities have emerged as promising candidates, reflecting PD-related dysfunction in ocular motor control pathways.
Design/Methods:
Eighty-one patients with idiopathic PD were enrolled in Zurich (Switzerland) and Exeter (UK). Each participant underwent two visits, including standardized ocular motor testing with the VR-based medical eye tracker and corresponding manual examination. Data with poor tracking quality or excessive signal loss were excluded. The present analysis included 132 PD examinations and 148 healthy control examinations from an independent dataset. Each PD patient was matched to one control of similar age and sex using a globally optimal matching algorithm. For each parameter, standardized mean differences (SMDs) and Welch’s tests were computed. A supervised machine learning model was trained to classify PD versus controls based on the most discriminative parameters. Site-specific analyses were conducted for Zurich and Exeter.
Results:
Parameters showing the largest PD–control differences included saccades (accuracy, velocity, main sequence), vergence, ocular alignment, and pupil dynamics (constriction/dilation velocity and latency). The machine learning model achieved strong discrimination: Zurich—AUC 0.96 (F1: 0.92, sensitivity: 0.96, specificity: 0.89); Exeter—AUC 0.91 (F1: 0.84, sensitivity: 0.89, specificity: 0.79).
Conclusions:
Preliminary findings demonstrate that quantitative ocular motor and pupillary metrics can robustly distinguish PD from healthy individuals. Next steps include expanding the patient and control cohorts, merging datasets across sites, and correlating ocular motor parameters with clinical measures to refine diagnostic performance.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.