DS Ph.D. Qualifier Presentation | Isaac Zhao | Tuesday, Dec. 12 @ Noon | Interpretable Sleep EEG Feature Discovery using a Deep Variational Autoencoder

DATA SCIENCE 

Ph.D. Qualifier Presentation

Isaac Zhao

 Tuesday, December 12th | 12:00 PM EST

Zoom: https://wpi.zoom.us/j/6074707249 

Committee: 

Prof. Carolina Ruiz, Dept. of Computer Science and Data Science Program, WPI

Prof. Sergio Alvarez, Dept. of Computer Science, Boston College

Prof. Fabricio Murai, Dept. of Computer Science and Data Science Program, WPI

Prof. Frank Zou, Dept. of Mathematical Science and Data Science Program, WPI

Title: 

Interpretable Sleep EEG Feature Discovery using a Deep Variational Autoencoder 

Abstract: 

Polysomnography, which includes electroencephalography (EEG) and other physiological signal types, is a key source of information in diagnosing sleep-related disorders. While deep learning models have achieved high performance in classifying EEG-based sleep stages, their underlying "black-box" lack of transparency limits the insight that they can provide from a sleep science perspective. Our research focuses on utilizing a Variational Autoencoder (VAE) deep learning framework to extract interpretable predictive features from EEG signals for sleep stage classification. We train separate VAE models for different EEG frequency bands and focus on shorter signal epochs than the standard 30s sleep epochs in order to improve feature learning. We use the latent traversal technique for interpretation of VAE internal features, in which latent variables are perturbed to observe the resulting changes in the decoder's output. Cyclical Kullback-Leibler (KL) annealing during VAE training and the use of nonstandard activation functions both lead to improvements in classification and reconstruction performance. Our work provides valuable insights into optimizing neural network architectures for EEG-based modeling.