fMRI is the neuroimage modality of choice when considering localized neurofeedback applications. However, the high costs and inflexibility of MRI setups limit their widespread application, motivating their transfer to EEG setups by reconstructing the BOLD-fMRI signal at the target regions using EEG only. Here, we systematically investigated the extent at which the BOLD-fMRI signal at the facial expressions processing network could be reconstructed from simultaneously recorded EEG signals. Features from both scalp and source spaces were extracted and used as predictors in a regression problem using random forests. We improved the accuracy of the state-of-the-art method from 20% to 53%.