Graph Neural Network Classification in EEG-Based Biometric Identification: Evaluation of Functional Connectivity Methods Using Time-Frequency Metric

The growing demand for secure identification has spurred interest in EEG-based biometric systems, which leverage unique brain activity patterns. Traditional univariate feature extraction methods, analyzing EEG channels independently, suffer from sensitivity to signal fluctuations. In contrast, bivariate features based on Functional Connectivity (FC) capture statistical interactions between brain regions, offering superior discriminability. However, conventional FC metrics like Correlation (COR) and Phase Locking Value (PLV) assume signal stationarity, failing to model EEG’s dynamic and directional nature. To address this, we introduce two novel time-frequency FC metrics: 1) Multi-band Coherency, which integrates features across delta, theta, alpha, and beta bands to enhance robustness and discriminability; and 2) wavelet coherency, addressing the non-stationarity of EEG signals, models phase lead-lag interactions and temporal variations. Integrated with Graph Convolutional Neural Networks (GCNNs), our approach leverages graph-structured FC data for superior classification. We propose a realistic experimental framework, recording training and testing signals in separate sessions under varying states, using only 19 EEG channels and single-trial signals. Despite reduced setup complexity, our GCNN achieves over 98% identification accuracy, comparable to CNN-based studies using 64 channels, with significantly lower computational cost and trainable variables reduced to less than 0.25 of those in a Convolutional Neural Network (CNN). In this study, self-recorded Auditory Evoked Potentials (AEPs), alongside public PhysioNet database, were used for their rapid acquisition, controllability, and ability to isolate individuals from external auditory stimuli. Statistical tests and connectivity visualizations validate our methods’ reliability, demonstrating advancements in EEG-based biometrics for secure and practical applications.