P 55 Functional connectivity and convolutional neural networks for automatic classification of EEG data

Abstract

It is well known that machine learning algorithms can support the classification of EEG data by using deep learning networks, especially convolutional neural networks [1-3]. In this paper, we propose a new method for automatic classification of movement disorders based on an open-access EEG dataset published by Anjun et al. in 2020 [4]. The dataset includes a total of 41 Parkinson's disease (PD) patients and 41 control subjects. Control participants were demographically matched for age and sex with PD patients and did not differ in any measurements of education or premorbid intelligence. Additionally, EEG recordings from OFF medication sessions for 27 PD patients were recorded in the practically defined OFF levodopa period (12 h after the last dose of dopaminergic medication). Resting state EEG recordings were gathered under both eyes-open and eyes-closed at a sampling rate of 500 Hz on a 64-channel system and eye blinks were removed [5]. In the first step of our machine learning process, a connectivity matrix is created using three different approaches based on Granger causality, Pearson correlation and Spearman correlation. In a second step, these matrices are fed into a convolutional neural network that is tuned using random search, hyperband and Bayesian optimization. We will show that this approach can provide very accurate classification results for the given EEG data sets. Comparison with the traditional method considering raw EEG data shows that our method is more accurate, highlighting the importance of network topology in describing brain data [6].