Interpreting the Filters in the First Layer of a Convolutional Neural Network for Sleep Stage Classification

Abstract

Sleep stage classification is the categorisation of Electroencephalogram (EEG) epoch into different sleep stages. Various supervised and unsupervised models have been developed for sleep stage classification. Emphasis of those models has been on classifying sleep stages using deep learning models such as the Convolutional Neural Network (CNN), however, very limited work exists on interpreting those CNN filters learned from EEG data in a supervised manner. This paper focuses on investigating and interpreting the output filters of the first CNN layer of the DeepSleepNet model, which is a model developed for automatic sleep stage scoring based on raw Single-Channel EEG. Experiments were carried out using a public benchmark dataset, namely the Sleep EDF Database. Spectral properties of both EEG epoch (input) and the learned filters obtained from the first CNN layer were compared. Results showed similar spectral properties between sleep EEG patterns and the learned filters which were obtained from the first CNN layer, and these findings suggest that ‘sleep stage’-defining EEG patterns are associated with certain learned CNN filters.