Decoding Human Attentive States from Spatial-temporal EEG Patches Using Transformers.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Decoding attention states from electroencephalogram (EEG) requires understanding channel spatial topology and temporal dynamics.
• Existing methods may not fully capture complex spatial-temporal EEG features for attention classification.

Purpose of the Study:

• To introduce EEG-PatchFormer, a transformer-based deep learning framework for EEG attention classification.
• To enhance Brain-Computer Interface (BCI) performance by improving EEG data decoding.

Main Methods:

• EEG-PatchFormer integrates Temporal CNN for frequency-based feature extraction and pointwise CNN for enhancement.
• Spatial and Temporal Patching modules organize features into spatial-temporal patches.
• A self-attention mechanism captures global spatial-temporal information across brain regions.

Main Results:

• EEG-PatchFormer demonstrated superior performance compared to existing benchmarks.
• The model achieved higher accuracy, area under the ROC curve (AUC), and macro-F1 score on a public cognitive attention dataset.

Conclusions:

• EEG-PatchFormer effectively learns crucial spatial-temporal information from EEG data for attention decoding.
• The proposed framework offers a promising advancement for BCI applications in attention state classification.