Microsaccade Dynamics During Visual Fixation as Markers for Parkinson's Disease: A Machine Learning Approach

Purpose: Microsaccades have long been of interest in vision and oculomotor research, but their potential as biomarkers in movement disorders such as Parkinson's disease (PD) remains underexplored. This study investigates the extent to which microsaccades carry information that can differentiate individuals with PD from a healthy population. Methods: We analyzed eye movements from 50 individuals with early-to-moderate stage PD and 43 healthy controls (HCs) recorded during a visual fixation task, without physically restricting head movements or interfering with the participants' natural behavior. To assess the predictive value of microsaccades for distinguishing PD from HCs, we train and evaluate a diverse set of machine learning classifiers on microsaccade features. Results: Microsaccades identified under these recording conditions show characteristics consistent with those reported in more controlled experimental settings. Microsaccades in the PD group occur more frequently, have larger amplitudes, and exhibit stronger horizontal bias than those in the HC group. Most classification models perform significantly better than chance. A polynomial support vector machine (SVM) achieves a classification accuracy of 77.4% on held-out test subjects, with balanced sensitivity (76.1%) and specificity (78.9%). Moreover, by discarding predictions with low confidence at the microsaccade level, performance can be further improved at the subject level. Conclusions: Microsaccades carry discriminative information capable of distinguishing individuals with PD from HCs, underscoring their potential value as early markers of motor dysfunction in PD. Translational Relevance: This study establishes a promising baseline against which future learning algorithms can be evaluated, with the long-term goal of developing clinically useful prediction models based on oculomotor behavior.