Pre-training of human sleep data improves efficiency and accuracy of mouse sleep stage classification by a transformer-based neural network

Abstract

We show that pre-training human sleep EEG data, but not other types of human EEG(electroencephalogram) data, improves both efficiency and accuracy of mouse sleep stage classification performed by a transformer-based neural network. First, a neural network with layered transformer architecture was designed to perform mouse sleep staging task. The training process involved a series of data augmentation methods tailored for electrophysiological data, which helped the network to robustly classify raw EEG traces into different sleep stages. Tested upon a publicly available mouse sleep dataset (Miladinovic et al., 2019), the classifier network showed comparable generalizability and accuracy to the state-of-the-art network models proposed for mouse sleep staging task. Especially, the layered architecture allowed neural network to consider surrounding signals to decide the sleep stage of a particular epoch, leading to improved performance. Next, the proposed network was used to test the transferability of information from human sleep EEG data to the sleep stage classification in mice. Pre-training with human sleep EEG data (Sleep-EDF; Kemp et al., 2000) has led the network to not only show higher accuracies at same epochs, but also show higher final accuracies. In contrast, pre-training with the same amount of motor imagery EEG data (EEG Motor Movement/Imagery Dataset; Schalk et al., 2004) rather deteriorated the performance of the network. These results suggest there are shared structures across the brain signals of humans and mice produced during sleep that allows inter-species transfer of particular information. The framework established in this study can be utilized to single out the similarities and peculiarities of sleep signals found in different species. It is to be tested whether information in electrophysiological signals from different domains can also be transferred across species.