Towards a Deeper Understanding of Sleep Stages through their Representation in the Latent Space of Variational Autoencoders

Abstract

Artificial neural networks show great success in sleep stage classification, with an accuracy comparable to human scoring. While their ability to learn from labelled electroencephalography (EEG) signals is widely researched, the underlying learning processes remain unexplored. Variational autoencoders can capture the underlying meaning of data by encoding it into a low-dimensional space. Regularizing this space furthermore enables the generation of realistic representations of data from latent space samples. We aimed to show that this model is able to generate realistic sleep EEG. In addition, the generated sequences from different areas of the latent space are shown to have inherent meaning. The current results show the potential of variational autoencoders in understanding sleep EEG data from the perspective of unsupervised machine learning.