Classifying REM Sleep Behavior Disorder through CNNs with Image-Based Representations of EEGs

ObjectiveRapid Eye Movement Sleep Behavior Disorder (RBD) is a parasomnia with a high conversion rate to -synucleinopathies such as Parkinsons Disease (PD), dementia with Lewy bodies (DLB), and multiple system atrophy (MSA). The objective of this paper is to classify RBD in patients through a convolutional neural network utilizing imagebased representations of electroencephalogram (EEG) channels. MethodsAnalysis was conducted on polysomnography data from 22 patients with RBD and 12 healthy controls acquired from the Cyclic Alternating Pattern (CAP) Sleep Database. EEG channels were split into four frequency bands (theta, alpha, beta, and gamma). Power spectrum density was calculated through a Fast Fourier Transformation (FFT) and converted into 2D images through bicubic interpolation. RBD classification was accomplished through a pre-trained VGG-16 CNN with layer weights being fine-tuned. ResultsThe model was successful in classifying RBD patients over non-RBD patients and achieved 97.92% accuracy over 40 epochs. Accuracy increased dramatically with increased data generated from FFT and interpolation (62.63% to 97.92%). ConclusionsThis project proposes a novel approach toward an automatic classification of RBD and highlights the importance of deep learning models in the field. The proposed transfer learning model outperforms state-of-the-art models and preliminarily accounts for the lack of computational resources in clinical spaces, thereby increasing the accessibility of automatic classification. SignificanceBy leveraging transfer learning and raw data, the results demonstrate that a similar model for the classification of RBD patients could easily be translated to the clinical atmosphere, drastically accelerating the classification pipeline. The proposed methods are also applicable to -synucleinopathies, including PD, DLB, and MSA.