Beyond the Surface: Revealing the Depths of Brain Activity by Predicting fMRI from EEG with Deep Learning
AO_SCPLOWBSTRACTC_SCPLOWElectroencephalography (EEG) and functional magnetic resonance imaging (fMRI) are the two most commonly used non-invasive methods for studying brain function, having different but complementary strengths: high temporal resolution of the former and high spatial resolution of the latter. Crucially, fMRI is vital for studying subcortical areas, as those are practically out of reach for conventional EEG. At the same time, EEG is cost-effective and, thus, often preferable to fMRI if comparable information could be extracted which is not the case then the deep subcortical brain activity is of interest. Here we explore the possibility of recovering subcortical hemodynamics from the non-invasively recorded scalp EEG signals. To this end, we have developed a lightweight EEG-to-fMRI neural network and using an extended and publicly available dataset with concurrently recorded EEG-fMRI data show that our model allows for the prediction of the detailed Blood Oxygenation Level Dependent (BOLD) activity of 7 bilaterally symmetric subcortical structures solely from multichannel EEG data. We report the performance significantly above chance and exceeding the scores achieved for a single subcortical structure and obtained on the proprietary datasets. In contrast to the studies focusing on a single subcortical region in our approach we were able to decode multichannel EEG into 14 + 4 region-specific variations of BOLD signals measured relative to their mean hemodynamic activity. The use of relative BOLD signals allowed us to exert control over the artificial inflation of decoding accuracy scores when the decoder predicts the common mode component that is likely to have a non-neuronal origin (heartbeat, movement, etc). Finally, we interpreted our model. The electrical activity of the sensorimotor cortex appeared to contribute most to the prediction of the subcortical hemodynamics. Also, the hemodynamics of the thalamus has the smallest delay with respect to the EEG signals. Both observations are physiologically plausible and ensure the potential reliability of the decoder. Taken together, these findings pave the road towards the creation of low-cost AI-powered EEG-based fMRI digital twin technology capable of tracking subcortical activity in an ecological setting. The technology, once mature, will find numerous applications from fundamental neuroscience through diagnostics to neurorehabilitation and affective neurointerfaces.12