Functional Connectivity Graph–Driven Imagined Speech Decoding for Assistive EEG-Based BCIs

Abstract

Imagined speech (IS) refers to the cognitive process of internally generating words without overt articulation and provides a promising communication pathway for individuals with severe neurological impairments. Electroencephalogram (EEG)-based brain–computer interface (BCI) systems enable automatic decoding of IS; however, reliable measurement and representation of EEG signals remain challenging due to their nonstationary nature and complex interchannel interactions. Most existing methods process EEG channels independently, neglecting functional connectivity patterns critical to speech-related neural dynamics. To address this limitation, our work proposes a sensor-aware graph signal processing (GSP)-based framework that models multichannel EEG measurements as graph-structured data, with electrodes represented as sensing nodes and weighted edges encoding functional relationships. Three graph construction strategies, namely correlation-based, Gaussian kernel-based, and Fundis-based weighting, are investigated to capture intersensor dependencies. Spectral Graph Wavelet Transform (SGWT) features are extracted to jointly characterize spatial connectivity, graph spectral information, and localized EEG signal variations at multiple graph scales, followed by statistical pooling for stability. The proposed framework is evaluated on a publicly available EEG dataset comprising four categories. The Fundis-based SGWT approach achieves classification accuracies of <inline-formula> <tex-math notation="LaTeX">$99.57{\,}\% {\,}\pm {\,}0.11{\,}\%$ </tex-math></inline-formula> for long words, <inline-formula> <tex-math notation="LaTeX">$99.60{\,}\% {\,}\pm {\,}0.15{\,}\%$ </tex-math></inline-formula> for short–long words, <inline-formula> <tex-math notation="LaTeX">$99.45{\,}\% {\,}\pm {\,}0.26{\,}\%$ </tex-math></inline-formula> for short words, and <inline-formula> <tex-math notation="LaTeX">$99.45{\,}\% {\,}\pm {\,}0.20{\,}\%$ </tex-math></inline-formula> for vowels, while maintaining computational efficiency with only 0.074 million parameters, 2.88 ms inference latency, and a compact model size of 4.35 MB. These results demonstrate the effectiveness of graph-based EEG modeling for real-time, low-power wearable BCI systems.