Laura Rose

PhD student, Kornum Lab

Title

Modeling EEG data using deep learning for automatic sleep stage classification

Abstract

In recent years, various new EEG deep learning models have been suggested for automatic sleep stage classification in mice. One of the major difficulties when modelling EEG is the large within-subject and between-subjects variation in EEG sleep patterns. The large variation raises the need for a model that is robust across different experimental paradigms. However, most of the existing classifiers have been trained on only a small and homogenous sample questioning the actual generalizability of these models.  For this reason, we want to conduct a validation study in which four of the state-of-the-art models are tested on external data from six different laboratories. We further intend to examine the effect of training the state-of-the-art models on a diverse sample and contrast it to an ensemble approach.

Despite the increasing numbers of proposed deep learning frameworks, none on them have yet succeeded in replacing the manual scoring in the laboratories. We believe that this is explained by the lack of interpretability which may reduce the trust to such models. A transformer model is a deep learning method that is inherently explainable through its attention scores. The attention scores enable us to examine how the model utilize properties of the EEG signal to determine its output. It is therefore further of interest to assess how the transformer model can be used to obtain a better understanding of sleep stage classification.