EEG-driven RNN Classification for Prognosis of Neurodegeneration in At-Risk Patients

Abstract

REM Behavior Disorder (RBD) is a serious risk factor for neurodegenerative diseases such as Parkinson’s disease (PD). We describe here a recurrent neural network (RNN) for classification of EEG data collected from RBD patients and healthy controls (HC) forming a balanced cohort of 118 subjects in which 50 % of the RBD patients eventually developed either PD or Lewy Body Dementia (LBD). In earlier work [1, 2], we implemented support vector machine classifiers (SVMs) using EEG mean spectral features to predict the course of disease in the dual HC vs. PD problem with an accuracy of 85 %. Although largely successful, this approach did not attempt to exploit the non-linear dynamic characteristics of EEG signals, which are believed to contain useful information. Here we describe an Echo State Network (ESN) classifier capable of processing the dynamic features of EEG power at different spectral bands. The inputs to the classifier are the time series of 1 second-averaged EEG power at several selected frequencies and channels. The performance of the ESN reaches 85 % test-set accuracy in the HC vs. PD problem using the same subset of channels and bands we selected in our prior work on this problem using SVMs.

This work partly supported by the Michael J. Fox Foundation within the project “Discovery of EEG biomarkers for Parkinson Disease and Lewy Body Dementia using advanced Machine Learning techniques”.