Automatic Classification of Sleep Stages from EEG Signals Using Riemannian Metrics and Transformer Networks

TL;DR

The capabilities of SPDTransNet, a Transformer-derived network designed to classify sleep stages from EEG data through timeseries of covariance matrices, are proved, particularly its adaptability to multi-dataset tasks, within the context of EEG sleep stage scoring.

Abstract

Purpose: In sleep medicine, assessing the evolution of a subject's sleep often involves the costly manual scoring of electroencephalographic (EEG) signals. In recent years, a number of Deep Learning approaches have been proposed to automate this process, mainly by extracting features from said signals. However, despite some promising developments in related problems, such as Brain-Computer Interfaces, analyses of the covariances between brain regions remain underutilized in sleep stage scoring.Methods: Expanding upon our previous work, we investigate the capabilities of SPDTransNet, a Transformer-derived network designed to classify sleep stages from EEG data through timeseries of covariance matrices. Furthermore, we present a novel way of integrating learned signal-wise features into said matrices without sacrificing their Symmetric Definite Positive (SPD) nature.Results: Through comparison with other State-of-the-Art models within a methodology optimized for class-wise performance, we achieve a level of performance at or beyond various State-of-the-Art models, both in single-dataset and - particularly - multi-dataset experiments.Conclusion: In this article, we prove the capabilities of our SPDTransNet model, particularly its adaptability to multi-dataset tasks, within the context of EEG sleep stage scoring - though it could easily be adapted to any classification task involving timeseries of covariance matrices.

Figures (5)

• Figure 1: Our data preparation pipeline, with $S$ = 30, $C$ = 7 and $n$ = 8.
• Figure 1: Epoch-wise feature extraction submodel of IITNet seo2020, adapted to extract augmentation features - with $n$ the number of EEG signals, $C$ the number of channels, and $k$ the size of each signal-wise feature vector.
• Figure 2: Estimation of the affine-invariant average of recording-wise enriched matrices in heatmap form, computed from unfiltered signals on the MASS-SS3 recording of index 42. From left to right, these correspond to DAW, MAW and WPA enrichment, respectively.
• Figure 3: Architecture of SPDTransNet, with $t$ = 3 feature tokens per epoch, $L$ = $2 \cdot \ell + 1$ the length of the input epoch sequence, and $\ell + 1$ the index of the central epoch. The "Matrix Computation & Tokenization" component is further developed in Figure \ref{['fig:signals_to_tokens']}.
• Figure 4: Our SP-MHA component. The computation of attention maps (small-dashed black rectangles) is identical to the original MHA, whereas the application of said attention maps to the tokens within the Value tensor (large-dashed red rectangle) has been modified to avoid any projection and subsequent concatenation.