Bandwidth-aware fusion of resting-state EEG-fMRI connectivity in cortical eigenmode space

We introduce a geometry-informed, bandwidth-aware model for fusing resting-state EEG and fMRI connectivity by treating each modality as a masked, physics-limited observation of a shared latent cortical connectivity structure. Both modalities are represented in a shared cortical Laplace-Beltrami (LB) eigenmode basis that provides an anatomically grounded ordering of spatial scales. In this coordinate system, fMRI constrains connectivity across a broad range of spatial scales but only at slow timescales, whereas EEG provides frequency-resolved information primarily for coarse spatial modes. We formalize this complementarity through a latent spatio-spectral connectivity object and estimate it using a low-rank factorization with shared eigenmode-network factors and subject- and frequency-specific nonnegative strengths. Estimation is driven by the regions supported by each modality (EEG: coarse spatial modes across broad frequencies; fMRI: broad spatial modes at low frequencies), while unobserved bands are inferred through shared low-rank structure and spectral smoothness regularization. In the MPI-LEMON cohort, the fused representation yields compact, interpretable subject features (fMRI network strengths and EEG oscillatory summaries) that support out-of-sample age prediction under nested cross-validation, and coherent multi-view interpretation through cortical maps, spectral profiles, and atlas-referenced system composition summaries. These results demonstrate that explicitly modeling modality-specific bandwidth gaps enables principled EEG-fMRI connectivity fusion and provides a practical route to multimodal network biomarkers for individual-differences research.