Classification of Sleep Stages Using Latent EEG Representations and Attention Mechanisms

• Grant number: 23/00255-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: April 01, 2023
• End date: March 31, 2024
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Raphael Yokoingawa de Camargo
• Grantee: Alexandre Janoni Bayerlein
• Host Institution: Centro de Matemática, Computação e Cognição (CMCC). Universidade Federal do ABC (UFABC). Ministério da Educação (Brasil). Santo André , SP, Brazil

Abstract

The process of classifying sleep stages is time-consuming and intensive and requires experts to analyze hours of EEG signals. The use of deep neural networks has been generating promising results, especially with convolutional architectures. These networks receive 30s windows as input and classify the sleep stage at that moment. However, contextual information, such as sleep stages at earlier and later times, is important for classification. Architectures like Transformers have attention mechanisms that allow the network to receive information from long EEG sequences and use those it deems relevant.A second problem in decoding EEG signals is that it has high dimensionality and is quite noisy, so large amounts of data are needed to train the neural network. Self-supervised learning techniques can be applied so that the neural network learns to extract relevant characteristics from the signal and generate a latent representation that can be later used by a classifier. This was done in an architecture called BENDR, which we will use to transform the EEG signal into low-dimensional latent representations.In this work, we will train a Transformer model to classify sleep stages. The model will receive as input latent representations of the EEG signal generated by the BENDR. Our hypothesis is that with low-dimensional latent representations, it will be possible to train the Transformer to use contextual information from periods of multiple minutes of sleep to generate more accurate classifications.