Early detection and severity evaluation of Parkinson's disease from eye movement data with the artificial intelligence algorithms

Human beings are more dependent on the sense of sight than other senses. This thesis aims to develop a system that will help in the early diagnosis and stage grading of Parkinson's disease (PD). Oculo-visual dysfunctions begin to be seen much earlier than the basic symptoms of the disease. However, the lack of an open access dataset in this field restricts the development of new methods and approaches. In this study, the signals on saccadic and pursuit eye movements of PD and healthy individuals were examined using the Videonystagmography (VNG) system, and these data were analyzed with machine and deep learning techniques. In this context, both spectral, fractal, statistical and temporal features of the signals were extracted with traditional methods and deep features from deep learning architectures were utilized. The performance of machine learning classifiers, such as Support Vector Machines (SVM), Naive Bayes (NB), k-Nearest Neighbors (k-NN), and Quadratic Discriminant Analysis (QDA) in complex classification tasks such as disease detection from eye movements and stage determination has been investigated. A two-stage feature selection method was applied, combining filtering and wrapper-based feature selection techniques, and it was determined that eliminating unnecessary features increased the classification performance of the model. Furthermore, evaluations with Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and various transfer learning architectures demonstrated the effectiveness of deep learning-based approaches in disease diagnosis. It was observed that saccadic eye movements showed higher success in the early-stage detection of the disease while pursuit movements were more effective in tracking disease progression. While the SVM algorithm come to the forefront in both eye movements in machine learning applications with the 91%-95% accuracy, long-short term memory showed a more successful performance with 72.73% in deep learning applications. Additionally, the hieghest success, with an accuracy rate of 79.55%, was achieved with GoogLeNet in applications with image representations of signals. This system based on eye movement analysis also has significant potential as an objective, easy-to-apply and efficient method in monitoring neurological diseases and treatment processes.