Electroencephalogram based attention detection becomes growing research filed in past years because his enormous application from education to clinical diagnosis. Exiting approaches to detect attention has often relied on features extracted from individual electrodes, such as frequency or entropy measures. In this study we examined graph-theoretical features derived from Coherence, Pearson correlation, and Mutual Information across theta, alpha, beta, and broadband bands in ten participants. Nine graph features capturing network integration and hub centrality were extracted per epoch, and false discovery rate correction identified 6 significant features in both theta, alpha bands and 9 in beta band, with notable effect sizes (Cohen’s d = 0.88–1.70). Theta networks showed small-world-like topology during Attention, alpha networks were more segregated during Inattention, and beta networks exhibited the richest differences. Subject-dependent classification reached 97.23% accuracy (Mutual information, Logistic Regression, beta band; F1-score = 0.970, Area under curve = 0.998), while leave one subject out confirmed broadband MI features as the most consistent (87.91% ± 6.34%; F1 = 0.893, AUC = 0.927), though higher inter-subject variability highlights instability was noted, which is expected in this small group of subjects. These results suggest that combining linear and information-theoretic connectivity measures captures complementary aspects of attentional networks, with broadband MI features outperforming traditional electrode-level approaches and offering a promising approach for EEG-based cognitive monitoring.