Brainwide fMRI patterns predict variations in EEG rhythms across sleep and wakefulness
Poster No:
2564
Submission Type:
Abstract Submission
Authors:
Leandro Jacob1, Sydney Bailes1, Stephanie Williams1, Carsen Stringer2, Laura Lewis1
Institutions:
1Massachusetts Institute of Technology, Cambridge, MA, 2Janelia, Ashburn, VA
First Author:
Co-Author(s):
Introduction:
Arousal-related EEG oscillations are associated with major changes in cognition, but identifying brainwide dynamics underlying them is challenging, as EEG suffers from low spatial resolution. EEG can be simultaneously acquired with fMRI-which has higher spatial resolution and can image deeper brain regions-but analyzing these multimodal data is challenging. Prior work focused on univariate mass correlations to identify individual fMRI regions coupled to EEG oscillations, but this approach cannot detect network-scale patterns. Here, we aimed to identify brainwide fMRI dynamics that underlie variations in neural oscillations. We used machine learning to find fMRI patterns that predict variations in EEG alpha (8-12Hz) and delta (1-4Hz) power during fluctuations in vigilance.
Methods:
We acquired simultaneous EEG and fast fMRI data in a 3T scanner (TR=378ms; voxel size=2.5mm isotropic) in subjects who spontaneously fell asleep (n=21 total, n=16 with alpha signals). We trained linear machine learning models to predict single timepoints of EEG power (5s window size) from 60 TRs of fMRI. fMRI data was parcellated in 84 regions—31 bilateral cortical, 7 bilateral subcortical, and 8 non-grey matter regions (white matter and ventricles). Models were iteratively trained on all subjects but one, and performance (correlation between predictions and ground truth) was assessed on held-out subjects. As control, we circularly shifted the fMRI by 2000 TRs in relation to the EEG (breaking the relationship between the modalities) prior to training and validation.
Results:
We successfully predicted EEG alpha and delta power from fMRI dynamics, and distinct fMRI regions carried information for distinct EEG bands (Fig 1). Alpha power was only predicted by gray matter regions, while delta was also predicted by white matter and ventricles, likely owing to the known coupling between delta and cerebrospinal fluid flow1. Demonstrating that the model can identify fine-grained information, EEG was also differentially predicted by individual bilateral fMRI regions, with alpha significantly predicted by arousal-controlling subcortical regions and V1 (replicating previous literature2–5). Delta was only significantly predicted by the putamen in this single-region analysis, suggesting that delta power requires a higher number of regions for successful predictions.
We thus investigated the spatial scales of alpha and delta by iteratively dropping low-weights regions (i.e. regions with redundant information) from the models (Fig 2A-B). Performance peaked at a much higher number for delta than for alpha, indicating that fMRI information predictive of delta is distributed widely (primarily through the cortex) while alpha is predicted by a much smaller set of subcortical and cortical regions. A clustering analysis showed that fMRI information predictive of delta is not only diffuse but not separable into distinct networks (Fig 2D). Alpha information, however, was clustered in two networks (Fig 2C): one primarily representing visual cortex and regions associated with higher-order cognition, and another highlighted by key arousal-controlling subcortical regions (Fig 2E-F). The regions in each network share the same predictive information, evidenced by worse within-cluster performance (mean r=-0.0163), while regions in different networks display distinct unique information, evidenced by improved between-cluster performance (mean r=0.0265; p<.001; Fig 2F-G).
We thus investigated the spatial scales of alpha and delta by iteratively dropping low-weights regions (i.e. regions with redundant information) from the models (Fig 2A-B). Performance peaked at a much higher number for delta than for alpha, indicating that fMRI information predictive of delta is distributed widely (primarily through the cortex) while alpha is predicted by a much smaller set of subcortical and cortical regions. A clustering analysis showed that fMRI information predictive of delta is not only diffuse but not separable into distinct networks (Fig 2D). Alpha information, however, was clustered in two networks (Fig 2C): one primarily representing visual cortex and regions associated with higher-order cognition, and another highlighted by key arousal-controlling subcortical regions (Fig 2E-F). The regions in each network share the same predictive information, evidenced by worse within-cluster performance (mean r=-0.0163), while regions in different networks display distinct unique information, evidenced by improved between-cluster performance (mean r=0.0265; p<.001; Fig 2F-G).
Conclusions:
Our work identified the fMRI network correlates of EEG delta and alpha across sleep and wakefulness, going beyond mass univariate relationships by extending a predictive approach to continuous cross-modal data. We found that alpha variations can be predicted by a small number of fMRI regions and that the fMRI information is divided in two networks-one represented by the visual system, and the other by arousal-controlling subcortical regions-while delta variations are predicted by a large and unified array of primarily cortical regions.
Modeling and Analysis Methods:
Classification and Predictive Modeling
Task-Independent and Resting-State Analysis
Novel Imaging Acquisition Methods:
BOLD fMRI
Multi-Modal Imaging 2
Perception, Attention and Motor Behavior:
Sleep and Wakefulness 1
Keywords:
Computational Neuroscience
Electroencephaolography (EEG)
FUNCTIONAL MRI
Machine Learning
NORMAL HUMAN
Sleep
1|2Indicates the priority used for review
Provide references using author date format
2. de Munck, J. C. et al. (2007). The hemodynamic response of the alpha rhythm: An EEG/fMRI study. Neuroimage 35, 1142–1151
3. Feige, B. et al. (2005). Cortical and subcortical correlates of electroencephalographic alpha rhythm modulation. J. Neurophysiol. 93, 2864–2872
4. Goldman, R. et al. (2002). Simultaneous EEG and fMRI of the alpha rhythm. Neuroreport 2487–2492
5. Tyvaert, L. et al. (2008). Effects of fluctuating physiological rhythms during prolonged EEG-fMRI studies. Clin. Neurophysiol. 119, 2762–2774