Brainwide hemodynamics predict EEG neural rhythms across sleep and wakefulness in humans

Abstract

The brain exhibits rich oscillatory dynamics that play critical roles in vigilance and cognition, such as the neural rhythms that define sleep. These rhythms continuously fluctuate, signaling major changes in vigilance, but the widespread brain dynamics underlying these oscillations are difficult to investigate. Using simultaneous EEG and fast fMRI in humans who fell asleep inside the scanner, we developed a machine learning approach to investigate which fMRI regions and networks predict fluctuations in neural rhythms. We demonstrated that the rise and fall of alpha (8-12 Hz) and delta (1-4 Hz) power-two canonical EEG bands critically involved with cognition and vigilance-can be predicted from fMRI data in subjects that were not present in the training set. This approach also identified predictive information in individual brain regions across the cortex and subcortex. Finally, we developed an approach to identify shared and unique predictive information, and found that information about alpha rhythms was highly separable in two networks linked to arousal and visual systems. Conversely, delta rhythms were diffusely represented on a large spatial scale primarily across the cortex. These results demonstrate that EEG rhythms can be predicted from fMRI data, identify large-scale network patterns that underlie alpha and delta rhythms, and establish a novel framework for investigating multimodal brain dynamics.

Fig 1. Machine learning can predict dynamic fluctuations in neural rhythms using brainwide fast fMRI timeseries.

a. Example of occipital alpha and delta EEG power fluctuating in one subject. b. Sleep stage distribution shows subjects drifted between wake and sleep (primarily N1 and N2). c. EEG alpha and delta power were separately predicted from simultaneous fMRI data, collected while subjects rested inside the scanner with their eyes closed. fMRI data was parcellated into 84 cortical, subcortical, and non-gray matter regions. Each point in the EEG power series, interpolated to match TR times, was predicted by 60 TRs of parcellated fMRI data. d. Correlation between alpha predictions (on held-out subjects) and ground truth, demonstrating that predictions using brainwide fMRI data are significantly better than the control condition, in which the fMRI data was shuffled within each subject run. *** p < 0.001, two-tailed paired t-test. Gray lines indicate held-out subject prediction performance. Error bars are SEM; N = 24. e. Examples of alpha prediction in individual subjects, showing tracking of short- and long-timescale alpha power fluctuations. f. Delta prediction performance, across the same conditions as panel d, shows significantly above-chance delta prediction. N = 32. g. Examples of delta prediction in individual subjects.

Fig 2. Alpha and delta rhythms show distinct relationships to cortical, subcortical, and non-neural fMRI signals.

a. Models were trained and tested under three additional conditions: cortical regions only, subcortical regions only, and non-gray matter regions only (ventricles and global white matter). Performance was compared to models trained using all regions, and under a control condition using all regions but with shuffled fMRI data, breaking the true relationship between EEG and fMRI. b. Alpha power can be predicted by cortical and subcortical regions, but not by non-gray matter regions. Circles show mean correlation between alpha power ground truth and model predictions (on held-out subjects). * p < 0.05; ** p < 0.01; *** p < 0.001; Tukey’s HSD. Error bars are SEM. Gray lines indicate individual held-out subject prediction performance. N = 24. b. Delta power can not only be predicted by cortical and subcortical regions, but also by non-gray matter regions on their own. N = 32.

Fig 3. Prediction performance maps show which regions individually carry significant information about alpha and delta rhythms.

a. Prediction performance (mean correlation between alpha predictions on held-out subjects and ground truth) when model was trained on single bilateral regions to predict alpha power. Regions shown in color yielded alpha predictions that were significantly better than control (non-significant regions are shown in grayscale). See Fig E in S1 Text for a list of correlation performance values. N = 18. b. Prediction performance for delta. Model was trained on single bilateral regions combined with non-gray matter areas to predict delta power, in order to account for the delta-predictive power of the non-gray matter (see Fig 2). Regions shown in color yielded predictions that were significantly better than non-gray matter areas alone. See Fig F in S1 Text for a list of correlation performance values. N = 32.

Fig 4. Clustering analysis of pairwise performance reveals two distinct networks of alpha-predictive fMRI information, while delta-predictive information is not separable in clusters.

a. Models were trained on every possible pair of gray matter regions, and correlation performance was calculated on held-out subjects. Pairwise performance benefits were calculated as the performance gain (or loss) of a pair in relation to the maximum performance of its individual components; see Fig 3 for individual performance values. Performance benefits were then clustered and evaluated with the gap statistic (Fig H in S1 Text) to determine the optimal clustering solution. b. Identified clusters and all pairwise performance benefits for alpha-predictive information. When regions from the blue and red cluster were paired together, they consistently obtained performance benefits not found in any other cluster match. N = 18. c. The clustering analysis revealed that delta did not possess a suitable clustering solution, as evidenced by no systematic pairwise performance benefits following a forced 3-cluster fitting. N = 32.