Lateralization in scalp EEG brain connectivity during hand motor imagery can improve task classification for...

This study evaluates brain connectivity reorganization during motor imagery (MI) tasks and assesses the predictive value of EEG-based functional connectivity measures for MI classification compared to µ-band (8-13 Hz) power spectrum of selected EEG channels, which are commonly used in MI decoders. We analyzed left- and right-hand MI EEG data from the BCI Competition IV 2a (BCI-IV-2a) and PhysioNet Motor Imagery (PHYS-MI) datasets. Phase Locking Value (PLV), cross-correlation (CC), weighted Phase Lag Index (wPLI), and Granger causality (GC) were evaluated as connectivity measures, and their decoding performance was compared against µ-band power features using Random Forest classifiers. Feature importance and graph-theoretical metrics were also used to examine node relevance, edge contributions, and global network topology across MI conditions. We found that PLV yields the most reliable MI decoding performance across both datasets, with accuracy comparable to power (65.3 ± 11.0% vs. 61.3 ± 11.0% and 58.4 ± 9.9% vs. 58.6 ± 15.7%, mean ± std. dev. across subjects for BCI-IV-2a and PHYS-MI, respectively). Moderate correlation (R ² = 0.62 and 0.40 for BCI-IV-2a and PHYS-MI, respectively) was found between the mean difference in PageRank centrality of the nodes of the PLV-based network in left- vs. right-hand MI and the Gini importance score of the single-channel power values. Also, while the PLV-based network topology remained stable over time, a small set of connections (7.8 ± 4.5% and 3.1 ± 2.5% of edges) lateralized to the hemisphere contralateral to the movement altered considerably and enhanced classification accuracy by 6.7 ± 5.6% and 16.3 ± 7.5% across subjects. These findings suggest that MI primarily modulates a limited number of task-specific functional connections. Rather than replacing established power-based approaches, connectivity measures provide complementary, network-level insight into how MI-related information is organized, which may inform interpretable feature selection and the design of future brain-computer interface models.