A self-supervised deep Riemannian representation to classify parkinsonian fixational patterns

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Parkinson's disease (PD) is a prevalent neurodegenerative disorder lacking definitive biomarkers for early detection or monitoring.
• Current diagnostic methods for PD often rely on subjective assessments or invasive procedures.
• Existing computational approaches for PD diagnosis typically require large labeled datasets and are prone to expert bias.

Purpose of the Study:

• To develop a self-supervised computational framework for identifying Parkinson's disease (PD) using oculomotor fixation patterns.
• To overcome limitations of current diagnostic methods, including invasive protocols and reliance on supervised learning with labeled data.
• To create a robust and unbiased method for detecting PD-related oculomotor abnormalities.

Main Methods:

• A self-supervised deep representation architecture was designed, utilizing Riemannian geometry.
• Deep convolutional features were extracted from oculomotor fixation video slices and encoded into compact Symmetric Positive Definite (SPD) matrices.
• A Riemannian encoder-decoder model was employed to learn geometric patterns from SPD embeddings without supervision.

Main Results:

• The proposed architecture successfully learned discriminative oculomotor fixation patterns without requiring diagnostic labels.
• In a binary classification task, the Riemannian representation achieved 95.6% average accuracy and 99% Area Under the Curve (AUC).
• The method demonstrated significant capability in distinguishing Parkinsonian patterns from healthy controls.

Conclusions:

• Self-supervised Riemannian deep representation offers a powerful, non-invasive approach for PD detection through oculomotor analysis.
• This method provides a sensitive biomarker for PD, potentially improving early diagnosis and patient management.
• The architecture effectively captures geometric patterns in oculomotor data, overcoming limitations of supervised learning and invasive methods.