Temporal Sequences of EEG Covariance Matrices for Automated Sleep Stage Scoring with Attention Mechanisms

Abstract

Electroencephalographic (EEG) data is commonly used in sleep medicine. It consists of a number of cerebral electrical signals measured from various brain locations, subdivided into segments that must be manually scored to reflect their sleep stage. These past few years, multiple implementations aimed at an automation of this scoring process have been attempted, with promising results, although they are not yet accurate enough with respect to each sleep stage to see clinical use. Our approach relies on the information contained within the covariations between multiple EEG signals. This is done through temporal sequences of covariance matrices, analyzed through attention mechanisms at both the intra- and inter-epoch levels. Evaluation performed on a standard dataset using an improved methodological framework show that our approach obtains balanced results over all classes, this balancing being characterized by a better MF1 score than the State of the Art.

This work has been co-funded by the Normandy Region and the French National Research Agency (ANR) through a HAISCoDe Ph.D. grant. It was granted access to the HPC resources of IDRIS under the allocation 2022-AD010613618 made by GENCI, and to the computing resources of CRIANN (Normandy, France).

Appendix

The hyperparameters corresponding to the best version of our model are:

Side vectors: PSD; $\alpha$: 99.53; intra-epoch encoder: 5 sublayers, 15 attention heads, fully connected components of size 1024, dropout of $6.2 \times 10^{-5}$; intra-epoch encoder: 6 sublayers, 5 attention heads, fully connected components of size 256, dropout of $8.1 \times 10^{-3}$; final fully connected layers: of size 2048, dropout of $1.4 \times 10^{-3}$; learning rate ($\lambda$): $4.9 \times 10^{-5}$, $\gamma _\lambda$ at 0.94; weight decay at $1.76 \times 10^{-6}$.

Many thanks to Huy Phan [11,12,13,14] for answering all our questions.