Enhancing brain activity mapping through BOLD-fMRI and MEG data fusion using explainable machine learning.

Motivation: Both magnetoencephalography (MEG) and functional MRI (fMRI) map brain activity, but with complementary spatial and temporal resolutions. Goal(s): Exploiting MEG and fMRI complementarity to enhance MEG spatial resolution via transfer learning from fMRI and explainable machine learning (xML). Approach: Sixteen participants underwent fMRI and MEG scans while watching the same movie clip in both. Tree-based ensemble learning models were trained to predict MEG from fMRI, enabling voxel-level super-resolution. Results: The trained model achieved high R² (0.93) on a test subset and outperformed naïve interpolation. Upsampled MEG maps showed greater specificity than interpolated images, suggesting promising applications for functional super-resolution. Impact: Our proof-of-concept approach demonstrates a strong predictive relationship between fMRI and MEG data. Using explainable machine learning, we generated high-resolution MEG maps from fMRI inputs, and provided data-driven insights into their relationship, opening new avenues for investigating neurovascular coupling.