Objective: To develop and evaluate a deep learning architecture (AntiSaccadeNet) that utilizes continuous antisaccade eye movement time-series data to differentiate between Parkinson’s disease (PD) and Progressive Supranuclear Palsy (PSP).
Background: Antisaccade eye movements engage complex neural circuits and offer potential for sensitive diagnostics in neurodegenerative disorders. Traditional methods rely on discrete metrics, potentially missing subtle temporal patterns diagnostic of different parkinsonian syndromes.
Method: Data were collected from 124 participants (individuals with de novo PD, medicated PD, PSP, and age-matched healthy controls), yielding over 11,000 antisaccade sequences. AntiSaccadeNet employs multi-resolution convolutional neural networks to process continuous time-series data of eye position, velocity, and acceleration. The architecture mirrors the hierarchical processing streams of neural circuits involved in antisaccade generation. Gradient-weighted Class Activation Mapping analysis identified influential temporal regions in classification decisions.
Results: AntiSaccadeNet achieved area under the receiver operating characteristic curves of 0.91-0.93 in distinguishing PSP from PD and healthy controls, significantly outperforming traditional feature-based approaches. The model identified distinct attribution patterns across diagnostic groups that align with known pathophysiology.
Conclusion: This calibration-free approach demonstrates the potential of deep learning models to capture rich temporal patterns in complex motor tasks like antisaccades. AntiSaccadeNet provides a more sensitive and specific diagnostic tool for parkinsonian disorders by processing continuous time-series data rather than discrete metrics. This represents a convergence of advanced computational methods with neuroscientific understanding, offering promise for early differential diagnosis in neurodegenerative diseases.
To cite this abstract in AMA style:
S. Patel, C. Antoniades, J. Fitzgerald, O. Bredemeyer. AntiSaccadeNet: A Deep Learning Framework for Differentiating Parkinsonian Disorders through Antisaccade Time Series Classification [abstract]. Mov Disord. 2025; 40 (suppl 1). https://www.mdsabstracts.org/abstract/antisaccadenet-a-deep-learning-framework-for-differentiating-parkinsonian-disorders-through-antisaccade-time-series-classification/. Accessed October 5, 2025.« Back to 2025 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/antisaccadenet-a-deep-learning-framework-for-differentiating-parkinsonian-disorders-through-antisaccade-time-series-classification/