Detection and Stage Classification of Parkinson's Disease with LSTM Networks Leveraging Eye Movement Data

Abstract

Parkinson's Disease (PD) is one of the progressive neurodegenerative disorders that affects motor control mechanism of patients. Abnormalities in smooth pursuit and saccadic eye movements can be used as early indicators of PD. This study explores the use of these oculomotor movement patterns as biomarkers for detecting the disease and assessing its severity by deep learning approaches. Long Short-Term Memory (LSTM) networks are especially proficient in capturing temporal relationships within time-series data. This property makes them ideal for examining dynamic recordings of eye movements. By employing LSTM neural network architecture, we first aim to classify individuals as either healthy or affected by PD, and in a further step, categorize the patients based on their disease stage. For this purpose, the oculomotor signals of 224 participants (112 healthy controls and 112 PD) were collected with a videonystagmography (VNG) system. A promising patient-healthy classification performance was achieved with 72.73% and 68.18% accuracies and 0.77 and 0.66 F-score values for saccadic and pursuit movements, respectively. On the other hand, pursuit movements come to the forefront in disease staging problem with the 73.0% accuracy and 0.84 F-score. As a result, analyzing saccadic eye movements would be more successful in systems to be developed for detecting the disease, while examining pursuit movements would provide higher success in systems to follow the progression of the disease. © 2025 Division of Signal Processing and Electronic Syste.