EEG-DATNet: Dual Attention Transformer for EEG Classification

Authors:
Preprint
Publisher website
Google Scholar
Abstract
EEG-based brain-computer interfaces remain constrained by signal non-stationarity, noise, and high dimensionality, limiting reliable decoding of motor imagery and cognitive states. EEG-DATNet is a hybrid architecture addressing these challenges through four components: multi-scale temporal convolutions spanning 60-260,ms capturing oscillatory content across canonical frequency bands; depthwise spatial convolutions encoding inter-electrode functional dependencies; a dual statistical pooling module pairing mean-activity and local-variance streams, the latter targeting event-related desynchronization patterns inaccessible to average pooling; and a spiking activation function trained via surrogate gradients yielding approximately 60% output sparsity, supporting energy-efficient inference. These streams feed a dual-attention transformer encoder applying independent temporal and spatial attention.On BCI Competition IV-2a (4-class motor imagery), EEG-DATNet achieves a mean accuracy of 82.84% (kappa = 0.77), reaching 96.82% (kappa = 0.95) for the best subject. On IV-2b (2-class), mean accuracy is 89.03%, outperforming EEG-Conformer (87.27%), EEGNet (85.04%), and FBCNet (81.95%). On the Thinking Out Loud dataset for inner speech decoding, EEG-DATNet achieves 43.68% using only 22 channels, versus 36.1% for BiLSTM and 29.67% for EEGNet. Ablation experiments confirm that dual attention is structurally necessary, with removal reducing accuracy by over 30 percentage points for certain subjects. Under Gaussian noise simulating motion artefacts, accuracy degrades by only 2.1%, versus 3.0% for EEGNet and 2.5% for EEG-Conformer. Preliminary estimates place inference at approximately 0.47, GFLOPs and 3.2 ms latency, and fine-tuning with 20% subject-specific data retains the majority of peak accuracy, suggesting practical viability in data-scarce clinical settings.

This publication has 18 references indexed in Scilit: