همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

DC-GATT: Dynamic Connectivity Graph Attention Transformer for Multi-Task EEG Classification

نویسندگان :

Mohammad Ahadzadeh¹ Pouya Taghipour Langrodi² Arian Afshar³ Golnaz Baghdadi⁴

1- دانشگاه صنعتی امیرکبیر 2- دانشگاه صنعتی امیرکبیر 3- دانشگاه صنعتی امیرکبیر 4- دانشگاه صنعتی امیرکبیر

کلمات کلیدی :

EEG،Dynamic functional connectivity،Graph attention network،Transformer،Multi-task classification،Weighted phase-lag index،Brain–computer interface

چکیده :

The decoding of cognitive states from electroencephalography (EEG) across multiple tasks and sessions remains challenging due to overlapping neural signatures and time-varying functional interactions. This paper introduces the Dynamic Connectivity Graph Attention Transformer (DC-GATT), an end-to-end model that combines fine-grained dynamic functional connectivity with transformer-based temporal attention for multi-task EEG classification. Using the COG-BCI dataset (29 subjects, 3 sessions; 62 channels), we represent each 1.2 s epoch as a sequence of 26 overlapping 200 ms graph snapshots (40 ms step), derive session-level weighted phase-lag index (wPLI) connectivity, sparsify with k-nearest neighbors (k=22), and extract node features via a temporal 1D convolutional encoder. Spatial context is modeled with a two-layer GATv2 stack, temporal dependencies with a multi-head Transformer encoder, and hierarchical attention pooling produces a fixed epoch embedding classified by a two-layer MLP. DC-GATT achieves an average test accuracy of 85.3 ± 5.3%, outperforming a static GAT baseline (70.9% ± 5.3%), with macro-averaged sensitivity 85.4%, specificity 92.7%, and F1-score 85.5%; best fold accuracy reached 92.6%. We show that sub-second dynamic connectivity carries discriminative information and that the proposed hierarchical graph–transformer architecture substantially improves multi-task, cross-session cognitive state decoding. Main contributions: (1) a novel DC-GATT architecture integrating dynamic wPLI graphs and transformers; (2) a hierarchical dual-attention pooling scheme; (3) empirical demonstration of superior multi-task performance on COG-BCI.