Eye-tracking is an accessible and non-invasive technology that provides information about a subject's motor and cognitive abilities. As such, it has proven to be a valuable resource in the study of neurodegenerative diseases such as Parkinson's disease. Saccade experiments, in particular, have proven useful in the diagnosis and staging of Parkinson's disease. However, to date, no single eye-movement biomarker has been found to conclusively differentiate patients from healthy controls. In the present work, we investigate the use of state-of-the-art deep learning algorithms to perform Parkinson's disease classification using eye-tracking data from saccade experiments. In contrast to previous work, instead of using hand-crafted features from the saccades, we use raw $\sim1.5\,s$ long fixation intervals recorded during the preparatory phase before each trial. Using these short time series as input we implement two different classification models, InceptionTime and ROCKET. We find that the models are able to learn the classification task and generalize to unseen subjects. InceptionTime achieves $78\%$ accuracy, while ROCKET achieves $88\%$ accuracy. We also employ a novel method for pruning the ROCKET model to improve interpretability and generalizability, achieving an accuracy of $96\%$. Our results suggest that fixation data has low inter-subject variability and potentially carries useful information about brain cognitive and motor conditions, making it suitable for use with machine learning in the discovery of disease-relevant biomarkers.

Time Series Classification (TSC) is essential in many fields, such as medicine, environmental science and finance, enabling tasks like disease diagnosis, anomaly detection, and stock price analysis. Machine learning models for TSC like Recurrent Neural Networks and InceptionTime, while successful in numerous applications, can face scalability limitations due to intensive computational requirements. To address this, efficient models such as ROCKET and its derivatives have emerged, simplifying training and achieving state-of-the-art performance by utilizing a large number of randomly generated features from time series data. However, due to their random nature, most of the generated features are redundant or non-informative, adding unnecessary computational load and compromising generalization. Here, we introduce Sequential Feature Detachment (SFD) as a method to identify and prune these non-essential features. SFD uses model coefficients to estimate feature importance and, unlike previous algorithms, can handle large feature sets without the need for complex hyperparameter tuning. Testing on the UCR archive demonstrates that SFD can produce models with $10\%$ of the original features while improving the accuracy $0.2\%$ on the test set. We also present an end-to-end procedure for determining an optimal balance between the number of features and model accuracy, called Detach-ROCKET. When applied to the largest binary UCR dataset, Detach-ROCKET is able to improve test accuracy by $0.6\%$ while reducing the number of features by $98.9\%$. Thus, our proposed procedure is not only lightweight to train and effective in reducing model size and enhancing generalization, but its significant reduction in feature count also paves the way for feature interpretation.