A Deep Convolutional-Recurrent Neural Network Architecture for Parkinson’s Disease EEG Classification

Abstract

Parkinson's disease (PD), characterized by slowness of movement, tremor and rigidity, is one of the most prevalent neurodegenerative disorders. Recent studies have demonstrated that abnormal neural oscillations within and between multiple brain regions play a critical role in the motor symptoms through invasive neural recordings. Progressions have been also made in EEG studies to use features in cortical oscillations recorded non-invasively as a diagnostic tool for PD. However, it is still challenging to effectively use EEG recordings for PD diagnosis. In this work, we design a novel deep learning framework for PD EEG classification. Specifically, the convolutional neural network (CNN) and the recurrent neural network (RNN) with long shortterm memory (LSTM) cells are exploited in our framework. First, we design two 1D-CNN layers to derive features to represent spatial (topological) relationships across EEG channels. Then, we apply LSTM on the spatial features from the CNN to further improve its performance. Finally, we validate our model on the PD classification on resting EEG recorded from 20 PD and 21 healthy subjects. Our method achieves accuracy of 96.9%, precision of 100%, and recall of 93.4% for differentiating PD from healthy controls and outperforms the state-of-the-art PD EEG classification results in the deep learning literature.