Modelling dynamic functional brain networks during sleep

1660 

Abstract Submission 

Hongyu Qian¹, Mark Woolrich¹, Chetan Gohil²

¹University of Oxford, Oxford, Oxfordshire, ²University of Oxford, Oxford, Oxford

Hongyu Qian  
University of Oxford
Oxford, Oxfordshire

Mark Woolrich  
University of Oxford
Oxford, Oxfordshire

Chetan Gohil  
University of Oxford
Oxford, Oxford

One major task of understanding brain networks is building computational methods that automatically (via unsupervised learning) learn how the brain's dynamics are organised into recurrent activations of transient brain networks. Recent studies have revealed the brain networks during resting state. However, limited work has been done to investigate the dynamics of such brain networks during sleep. Researchers have proposed a new approach DyNeMo (Dynamic Network Modes) that uses a new generative model for electrophysiological data and a Bayesian framework to learn model parameters from the observed data. This approach has been proven to be effective in analysing both resting-state and task M/EEG data. This study therefore set out to assess the performance of DyNeMo on sleep EEG data. DyNeMo was able to identify reasonable networks in sensor space. Finally, it is concluded that DyNeMo is a reliable tool to analyse sleep EEG data and the study can now extend to source space.

Researchers have proposed a new method that builds on recent advances in deep learning to capture the rich spatio-temporal content of brain network dynamics that underpins cognition. The current approach models brain networks as a time-varying linear mixture of spatially distributed modes. The temporal evolution of this mixture is governed by a recurrent neural network (RNN), which enables the model to generate data with a rich temporal structure. This approach of Dynamic Network Modes called DyNeMo is proposed by Gohil et al. in "Mixtures of large-scale dynamic functional brain network modes".
Data from the research of Schreiner et al. was used in this study. Twenty participants (age: 20.75±0.35; 17 female) took part in two experimental sessions and we received data for nineteen subjects. In both sessions, they performed an episodic learning task, with memory performance being assessed before and after taking a 120-min nap. A Brain Products 64-channel EEG system was used to record electroencephalography (EEG) throughout the experiment. Impedances were kept below 10 kΩ. EEG signals were referenced online to electrode FCz and sampled at a rate of 1000 Hz. Furthermore, electromyogram (EMG) and electrocardiogram (ECG) were recorded for polysomnography. Sleep architecture was determined offline according to standard criteria by two independent raters.

In Figures 1, we show topographic power maps and PSDs of 8 modes inferred by DyNeMo when trained on the sleep dataset. A General Linear Model (GLM)-based approach was used on a spectrogram directly calculated from the data to calculate mode and region-specific PSDs.Mode 1 appears to be a low-power background mode, whereas modes 2-8 show high power with different frequency bands at different locations. Figure 3 shows more detailed topographic maps for each mode in different frequency bands. We can observe that mode 4 shows clear alpha activity at occipital regions and has the strongest beta activity across the modes. Mode 8 is the only mode with significant sigma power and the power is stronger near the F3 and F4 electrodes. Modes 2 and 3 show high power in the low-delta band at the centralposterior region and frontal region respectively, while mode 7 shows low-delta power near a specific electrode. Modes 2,3,4 and 8 parallel the characteristic patterns observed in typical sleep EEG results. This correlation supports the relevance and validity of the DyNeMo modes.
The classifier achieved an average accuracy of 68.2% and only made few blatant errors, for example, mislabelling N2 and N3 to wake or REM.This indicates their potential for furthering our understanding of dynamic functional brain networks during sleep.

This study has used DyNeMo to identify distinct modes that are broadly in line with previous EEG studies The application of DyNeMo to sleep EEG data reveals modes that could relate to sleep spindles and slow oscillation, suggesting that it remains a novel and complementary tool for studying sleep data.

EEG/MEG Modeling and Analysis ¹

Sleep and Wakefulness ²

Computational Neuroscience

Data analysis

Electroencephaolography (EEG)

Machine Learning

Open-Source Code

Sleep