[Paper Review] Decoding Human Attentive States from Spatial-temporal EEG Patches Using Transformers

Yi Ding, Joon Hei Lee+3 more|ArXiv.org|Feb 6, 2025

EEG and Brain-Computer Interfaces3 citations

TL;DR

EEG-PatchFormer is a transformer-based model that jointly learns spatial-temporal EEG features via spatial/temporal patches and patches them for a Transformer to classify attention states, achieving state-of-the-art results on a public cognitive attention dataset.

ABSTRACT

Learning the spatial topology of electroencephalogram (EEG) channels and their temporal dynamics is crucial for decoding attention states. This paper introduces EEG-PatchFormer, a transformer-based deep learning framework designed specifically for EEG attention classification in Brain-Computer Interface (BCI) applications. By integrating a Temporal CNN for frequency-based EEG feature extraction, a pointwise CNN for feature enhancement, and Spatial and Temporal Patching modules for organizing features into spatial-temporal patches, EEG-PatchFormer jointly learns spatial-temporal information from EEG data. Leveraging the global learning capabilities of the self-attention mechanism, it captures essential features across brain regions over time, thereby enhancing EEG data decoding performance. Demonstrating superior performance, EEG-PatchFormer surpasses existing benchmarks in accuracy, area under the ROC curve (AUC), and macro-F1 score on a public cognitive attention dataset. The code can be found via: https://github.com/yi-ding-cs/EEG-PatchFormer .

Motivation & Objective

• Motivate decoding of human attentive states from EEG by jointly modeling spatial and temporal dynamics.
• Introduce a patch-based Transformer architecture to capture cross-regional brain activity over time.
• Design modules to extract frequency-based features, enhance representations, and organize them into spatial-temporal patches for Transformer processing.
• Demonstrate superior performance against existing EEG attention classification methods on a public dataset.

Proposed method

• Temporal CNN extracts frequency-based temporal features from EEG signals.
• Feature Enhancement with a 1x1 convolution to refine and reduce feature dimensionality.
• Spatial Patching Module with local and global branches to generate spatial patches.
• Local spatial learner with brain-region grouped processing and channel-wise filtering.
• Global spatial learner with a final CNN over all channels.
• Temporal Patching Module creating overlapped temporal patches via sliding window and linear projection.
• Transformer encoder with 4 layers processes spatial-temporal tokens, followed by a fully connected classifier.

Experimental results

Research questions

• RQ1 Can a transformer-based model leveraging spatial-temporal EEG patches improve attention-state classification over existing methods?
• RQ2 Does joint learning of local/global spatial information and temporal dynamics via patching enhance decoding performance?
• RQ3 What is the impact of patch design choices (overlap, lengths) on classification metrics?
• RQ4 How does EEG-PatchFormer compare to baseline EEG architectures in ACC, AUC, and macro-F1 on a public attention dataset?

Key findings

• EEG-PatchFormer achieves the highest mean ACC of 75.63% across compared baselines.
• It attains the highest AUC of 85.06% among methods evaluated.
• It achieves the top macro-F1 score of 75.04%.
• Ablation shows that removing components (FEM, SPM, OTPM) degrades performance, especially without SPM.
• Optimal temporal patch length in TPM is 20 data points for best metrics.
• Overall, the model outperforms DGCNN, LGGNet, TSception, and ViT on ACC, AUC, and macro-F1.