Toward Enhanced Neural Decoding: A Framework for Reconstructing EEG Features from fMRI BOLD Signals

Simultaneous EEG-fMRI acquisition integrates the complementary strengths of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI), offering a more comprehensive understanding of brain dynamics. However, EEG data acquisition can sometimes be limited or compromised due to electrode issues, motion artifacts, or participant discomfort, underscoring the importance of methods to infer EEG spectral features from available fMRI recordings. In this study, we propose a deep learning framework that directly reconstructs EEG power spectral density (PSD) features from resting-state fMRI BOLD signals. By aggregating spatiotemporal information across 200 brain regions over consecutive time windows, our model effectively captures complex fMRI dynamics and maps them onto neural electrical activity. We evaluate our method on a simultaneous EEG-fMRI dataset comprising 24 healthy participants. PSD features are reconstructed across five canonical EEG frequency bands: Delta, Theta, Alpha, Beta, and Gamma. Experimental results demonstrate high intra-subject prediction performance, with average Pearson correlations exceeding 0.90 in the Gamma and Theta bands, and mean squared errors (MSE) as low as 0.0035. The highest band-specific correlation reaches 0.9886 in the Gamma band, highlighting a strong coupling between BOLD activity and high-frequency neural oscillations. Topographic analyses further confirm that the model accurately captures both subject-level and group-level EEG spatial patterns. To the best of our knowledge, this is the first study to directly decode EEG signals from fMRI signals, offering a pathway toward precise brain activity interpretation and advancing the integration of multimodal neuroimaging.