Estimation of fMRI-informed EEG models of motor imagery brain activity from simultaneous EEG-fMRI

During recent years, the integration between EEG and fMRI has been largely investigated due to their highly complementary characteristics, leading to different multimodal approaches recognised as promising new tools in different conditions, especially the post-stroke neurorehabilitation process, which strongly benefits from the recent advancements in BCI-NF. However, the use of fMRI has been deemed the main limitation due to its high cost, low portability and discomfort for the patient. This sparked the research for models completely based on the widely-available EEG which are capable of providing information about the simultaneous BOLD signal. These models, called fMRI-informed EEG models, are a promising technique which still needs to be explored in detail for the post-stroke neurorehabilitation. Within this scope, this work relies on a dataset of 15 healthy participants who underwent two sessions of simultaneous EEG-fMRI while performing a Motor Imagery task. A wide range of relevant EEG features are extracted and used as predictors in a distributed lag model which aims at reconstructing the simultaneous BOLD response either by shrinkage (Elastic Net) or feature ranking (mRMR) linear regression. A specific Cross Validation design is then applied to optimize the hyperparameters and test the models. The obtained results show that the Elastic Net model outperforms the mRMR one showing an easier optimization process and yielding a better average performance on the test set, measured as the Pearson correlation coefficient between the predicted fMRI and the true signal, respectively of r_EN = 0.17±0.12 and r_mRMR = 0.12±0.08. The findings of this work provide relevant insights about many confounds present in the fMRI-informed EEG field applied to Motor Imagery, and constitutes a starting point for future improvements achievable by inspecting different feature selection approaches and non linear regression methods.