Deep learning with EEG spectrograms in rapid eye movement behavior disorder

REM Behavior Disorder (RBD) is now recognized as the prodromal stage of -synucleinopathies such as Parkinsons disease (PD). In this paper, we describe deep learning models for diagnosis/prognosis derived from a few minutes of eyes-closed resting electroencephalography data (EEG) collected from idiopathic RBD patients (n=121) and healthy controls (HC, n=91). A few years after the EEG acquisition (4 {+/-} 2 years), a subset of the RBD patients eventually developed either PD (n=14) or Dementia with Lewy bodies (DLB, n=13), while the rest remained idiopathic. We describe first a simple convolutional neural network (DCNN) with a five-layer architecture combining filtering and pooling, which we train using stacked multi-channel EEG spectrograms. We treat the data as in audio or image classification problems where deep networks have proven highly successful by exploiting compositional and translationally invariant features in the data. For comparison, we study an even simpler deep recurrent neural network using three stacked Long Short Term Memory network (LSTM) cells or gated-recurrent unit (GRU) cells--with very similar results. The performance of these networks typically reaches 80% ({+/-}1%) classification accuracy in the balanced HC vs. PD-outcome classification problem. In particular, using data from a single EEG channel we obtain an area under the curve (AUC) of 87% ({+/-}1%) while avoiding spectral feature selection. The trained classifier can also be used to generate synthetic spectrograms to study what spectrogram features are relevant for classification, pointing to the presence of theta band bursts and a decrease of power in the alpha band in future PD or DLB patients compared to HCs. We conclude that deep networks may provide a key tool for the analysis of EEG dynamics even from relatively small datasets and enable the delivery of new biomarkers.