Exploring Internal Representations Learned by Autoencoders for Sleep EEG

Soota, Akshit

Etd

Exploring Internal Representations Learned by Autoencoders for Sleep EEG

Public Deposited

The current gold standard for describing sleep progression relies on visual classification by human technicians of 30-second physiological signal traces, particularly electroencephalograms (EEG), into sleep stages. Supervised learning using deep neural networks such as convolutional neural networks (CNNs) can deliver automated sleep staging classification performance that is on par with human technicians without the shortcomings of visually determined staging. Our work shifts the focus to self-supervised learning, using variational and denoising autoencoders. We aim to understand the internal representations learned by such neural networks over EEG input data in the form of time-frequency spectrograms through visualization and analysis of the learned features; we also note the potential for using symbolic descriptions of such features. Our approach can advance the understanding of sleep EEG signals and contribute to sleep medicine.

Creator