[Paper Review] Deep Feature Mining via Attention-based BiLSTM-GCN for Human Motor Imagery Recognition

Yimin Hou, Shuyue Jia+5 more|arXiv (Cornell University)|May 2, 2020

EEG and Brain-Computer Interfaces49 references32 citations

TL;DR

The paper introduces an Attention-based BiLSTM-GCN framework to decode EEG motor imagery with 0.4s samples, achieving state-of-the-art group-level and subject-level accuracies.

ABSTRACT

Recognition accuracy and response time are both critically essential ahead of building practical electroencephalography (EEG) based brain-computer interface (BCI). Recent approaches, however, have either compromised in the classification accuracy or responding time. This paper presents a novel deep learning approach designed towards remarkably accurate and responsive motor imagery (MI) recognition based on scalp EEG. Bidirectional Long Short-term Memory (BiLSTM) with the Attention mechanism manages to derive relevant features from raw EEG signals. The connected graph convolutional neural network (GCN) promotes the decoding performance by cooperating with the topological structure of features, which are estimated from the overall data. The 0.4-second detection framework has shown effective and efficient prediction based on individual and group-wise training, with 98.81% and 94.64% accuracy, respectively, which outperformed all the state-of-the-art studies. The introduced deep feature mining approach can precisely recognize human motion intents from raw EEG signals, which paves the road to translate the EEG based MI recognition to practical BCI systems.

Motivation & Objective

• Motivate fast and accurate EEG-based MI recognition for practical BCI by addressing inter-trial and inter-subject variability.
• Develop a deep learning pipeline that learns robust features from raw 64-channel EEG data.
• Leverage topological feature structure to enhance decoding via graph-based learning.
• Demonstrate robustness and repeatability through group-wise and subject-specific evaluations.

Proposed method

• Use 64-channel EEG signals sliced into 0.4-second segments (64x64) as input.
• Apply BiLSTM with an Attention mechanism to extract spatial-temporal features from raw EEG.
• Construct a graph representation of extracted features using Pearson, absolute Pearson, adjacency, and Laplacian matrices.
• Classify the graph-structured features with a six-layer Graph Convolutional Network (GCN) with pooling and softmax output.
• Train end-to-end in TensorFlow with dropout, batch normalization, and cross-entropy loss.
• Tune hyperparameters to maximize performance while preventing overfitting (e.g., 256 BiLSTM cells, 8 Attention size, 64 FC size).

Experimental results

Research questions

• RQ1 Can an Attention-based BiLSTM extract salient EEG MI features from short 0.4s segments more effectively than standard RNNs or CNNs?
• RQ2 Does incorporating a Graph Convolutional Network on top of extracted features improve MI classification by leveraging feature topology?
• RQ3 What are the group-wise and subject-specific performance gains when using this hybrid BiLSTM-GCN approach on MI-EEG data?

Key findings

• Group-wise max GAA of 94.64% with 20 subjects using 10-fold cross-validation.
• Subject-wise max GAA of 98.81% across individual subjects.
• Average group-wise performance (GAA) of 93.04% with high reliability metrics (Kappa 90.71%, Precision 93.02%, Recall 93.01%, F1 92.99%).
• Subject-level adaptation achieved an average GAA of 95.48% across 10 subjects (± metrics).
• The method demonstrates robustness to inter-subject variability and provides time-resolved predictions with 0.4s samples, outperforming several state-of-the-art approaches.