How much of the BOLD-fMRI signal can be approximated from simultaneous EEG data: relevance for the transfer and dissemination of neurofeedback interventions.

fMRI-based neurofeedback (NF) interventions represent the method of choice for the neuromodulation of localized brain areas. Although we have already validated an fMRI-NF protocol targeting the facial expressions processing network (FEPN), its dissemination is hampered by the economical and logistical constraints of fMRI-NF interventions, which may be however surpassed by transferring it to EEG setups, due to their low cost and portability. One of the major challenges of this procedure is then to reconstruct the BOLD-fMRI signal measured at the FEPN using only EEG signals. Because these types of approaches have been poorly explored so far, here we systematically investigated the extent at which the BOLD-fMRI signal recorded from the FEPN during a fMRI-NF protocol could be reconstructed from the simultaneously recorded EEG signal. Several features from both scalp and source spaces (the latter estimated using continuous EEG source imaging) were extracted and used as predictors in a regression problem using random forests. Furthermore, three different approaches to deal with the hemodynamic delay of the BOLD signal were tested. The resulting models were compared with the only approach already proposed in the literature that uses spectral features and considers different time delays. The combination of linear and non-linear features (particularly the largest Lyapunov exponent and entropy measures) projected into the source space, spatially filtered by independent component analysis (ICA) and convolved with multiple HRF functions peaking at different latencies, increases significantly the reconstruction accuracy (defined as the correlation between the measured and approximated BOLD signal) from 20% (direct comparison with the method used in the current literature) to 56%. With this pipeline, a more accurate reconstruction of the BOLD signal can be obtained, which will positively impact the transfer of fMRI-based neurofeedback interventions to EEG setups, and more importantly, their dissemination and efficacy in modulating the activity of the desired brain areas.