Automated Eye-Tracking for Parkinson's Disease Diagnosis: A Proof-of-Concept Cascade Classifier Study Establishing Clinical Validity

Abstract Background. Parkinson's disease (PD) is a progressive neurodegenerative disorder of increasing prevalence, with diagnostic accuracy of approximately 26% in early symptomatic patients. There is a need for accurate, non-invasive biomarkers to aid in disease diagnosis. Methods. This proof-of-concept study enrolled 90 participants (PD n = 30, other movement disorders [OM] n = 30, healthy controls [HC] n = 30) at a single institution. Participants completed two 10-minute eye-tracking sessions using the SaccadeDX 250 Hz binocular system. A two-level cascade classifier was fitted using elastic-net feature selection followed by logistic regression on the selected features, validated by 10-fold cross-validation. The cascade distinguished HC from movement disorders (Level 1) and PD from OM (Level 2), with the objective of establishing clinical validity that an eye-tracking signal correlates reliably with PD diagnosis. Results. Level 1 achieved an area under the curve (AUC) of 0.818 (95% CI: 0.71, 0.91), with a sensitivity of 83% and specificity of 63%. Level 2 achieved an AUC of 0.670 (95% CI: 0.52, 0.80), with a sensitivity of 68% and specificity of 63%. End-to-end PD detection achieved an AUC of 0.866 and an accuracy of 83.5%, meeting the prospectively specified accuracy threshold and the proof-of-concept AUC benchmark. Five adverse events were recorded (three cases of dizziness, one of nausea, and one of dry eyes); one participant withdrew from the study. Conclusions. Clinical validity is established: a reproducible eye-tracking signal for PD is detectable using a two-level cascade classifier. A multi-center confirmatory study is warranted before assessment of clinical utility.