Abnormal fixational eye movements are frequently observed in individuals with Parkinson’s disease (PD). While these involuntary movements have the potential to serve as objective indicators for early-stage PD screening, the specific role of microsaccades in this context has yet to be validated using computational methods. In this study, we combine explainable AI (XAI) techniques with end-to-end classification to evaluate the contribution of microsaccades in differentiating early-stage PD from healthy controls using data from a simple fixation task. A hybrid deep learning architecture fully utilizing temporal and spatial aspects of eye movements will be designed to analyze raw binocular eye position data, while different XAI methods such as integrated gradients will be used to identify the most influential samples in the model's decision-making process. Since our dataset suffers from the data scarcity, which is a common issue in medical application, generative algorithm will be employed to generate more synthesis data for training. Our preliminary experimental results show that it is feasible to distinguish early-stage PD based on raw eye tracking data. Importantly, visualizing integrated gradients across eye position over time indicates that microsaccades play a key role in classification, underscoring their potential as discriminative features for early-stage PD screening.