Simultaneous electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) acquisitions can provide more effective neurofeedback (NF) training due to their complementary temporal and spatial precision. However, MRI is expensive and can be draining for participants. Therefore, our goal is to reduce the reliance on MRI by developing a model that can predict fMRI NF scores from EEG signals alone, potentially eliminating the need for MRI. Yet, arbitrarily proposing a model architecture for such complex problems is challenging. So, in this study, we used a genetic algorithm to search for neural network architecture hyperparameters, specifically applied here to convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. The resulting architectures provided fMRI NF score predictions that, when combined with EEG NF scores, significantly matched the true bi-modal EEG-fMRI NF scores more closely than the EEG NF scores alone. This approach demonstrates the potential for enriching the EEG modality in a unimodal neurofeedback framework, thereby reducing the need for MRI. However, the predictions still lack precision. Therefore, this work thoroughly investigates the potential for enriching the EEG modality in a unimodal neurofeedback framework. Our code and models are available at https://gitlab.inria.fr/cpinte/prediction-of-fmri-neurofeedback-scores-from-eeg-signals.