Search Papers | Poster Sessions | All Posters
Poster B143 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Machine learning predicts neural rhythms from brainwide hemodynamics across vigilance states
Leandro Jacob1 (), Sydney Bailes1,2, Stephanie Williams1,2, Carsen Stringer3, Laura Lewis1,4; 1Massachusetts Institute of Technology, 2Boston University, 3HHMI Janelia Research Campus, 4Massachusetts General Hospital
Neurons often fire in synchrony, generating rhythms that support cognition and signal distinct brain states. These rhythms have been widely studied with EEG, but EEG’s low spatial resolution limits our ability to investigate the brainwide activity that underlies neural rhythms. fMRI can measure brainwide activity through hemodynamic signals, but identifying relationships between hemodynamics and electrophysiology is analytically challenging, particularly when trial averaging is not possible—such as in studies of spontaneous, naturally varying brain states. We developed a machine learning approach that predicts neural rhythms (EEG power in canonical frequency bands) from fast fMRI (<400 ms TR). Using two datasets of participants (n=21) drifting in and out of sleep, we show that neural rhythms can be predicted from brainwide fMRI dynamics in out-of-sample subjects, and that different patterns of fMRI regions predict alpha (8-12Hz) and delta (1-4Hz) EEG power. Alpha was primarily predicted by arousal-controlling subcortex and V1, while delta predictions relied on a large number of primarily cortical regions, with significant contributions from the putamen and non-gray matter components. Our results reveal the brainwide activity underlying key neural rhythms involved in cognition and arousal, and enables discovery of the large-scale dynamics linked to neural rhythms, with applications to diverse neuroscience questions.
Keywords: machine learning EEG fMRI sleep