Capturing Class Differences in EEG Connectivity with Class-discriminative Multivariate IMagInary Coherency (CMIMIC)

A major challenge in EEG/MEG analysis is to characterize task-dependent functional connectivity in a way that is statistically robust and spatially interpretable, without prior source modeling or region-of-interest selection. Comparing connectivity between conditions is statistically demanding due to the high dimensionality of connectivity estimates, and spatial interpretability is compromised when effects are fragmented across many tests. Here, we introduce Class-discriminative Multivariate ImagInary Coherency (CMIMIC), a multivariate method that maximizes differences in the imaginary part of coherency between two classes, enabling the robust identification of condition-specific functional interactions while minimizing spurious connectivity due to volume conduction. We evaluated CMIMIC using simulations and resting-state EEG data of young and older adults in eyes open and eyes closed conditions. In simulations, CMIMIC recovered both the temporal dependencies between sources and their underlying generating networks. On empirical data, the CMIMIC-based pipeline consistently outperformed those based on traditional imaginary coherence (ImCoh) across classifiers and conditions. All methods performed better with eyes open than closed, but CMIMIC showed only a modest performance drop, in contrast to substantial losses for ImCoh. Feature stability analyses revealed that CMIMIC identifies a core set of informative sources largely shared across conditions, whereas ImCoh features were more condition-specific. CMIMIC also provides spatially interpretable source configurations that align well with known neurophysiological effects, offering a reliable, interpretable, and condition-robust approach for extracting discriminative connectivity features from EEG data.