GACL-Net

GACL-Net: Hybrid Deep Learning Framework for Accurate Motor Imagery Classification in Stroke Rehabilitation

This repository contains the implementation of graph-attentive convolutional long short-term memory network (GACL-Net), a novel deep learning model designed for accurate motor imagery (MI) classification in stroke rehabilitation. GACL-Net integrates multi-scale convolutional layers, attention fusion mechanisms, graph convolutional networks, and bidirectional LSTMs to enhance classification robustness and generalization across stroke patients.

Project Structure

This repository includes scripts for feature extraction, feature selection, model definition, training, and statistical analysis:

1. Feature Extraction (FeaturesExtraction.py)

Extracts spatial, temporal, and spectral features from EEG signals, including:

• Alpha & Beta Band Power
• Hilbert Amplitude Envelope
• EEG Coherence
• Event-Related Desynchronization (ERD/ERS)
• Fractal Dimension & Lyapunov Exponent

2. Feature Selection (GA_Fetures_Selection.py)

Implements genetic algorithm (GA) for optimal feature subset selection, reducing model complexity while maintaining high accuracy.

3. Model Definition (GACL_Model.py)

Defines the GACL-Net architecture, including:

• Multi-scale convolutional block
• Attention fusion layer
• Graph convolutional layer
• Bidirectional LSTM with attention
• Hierarchical feature aggregation & dense layers

4. Model Training & Evaluation (Model_Training.py)

• Loads extracted features and applies data augmentation & normalization.
• Splits the dataset into training, validation, and test sets.
• Trains GACL-Net with cross-entropy loss & Adam optimizer.
• Evaluates accuracy, precision, recall, and F1-score.

5. Statistical Analysis (Statistical_Analysis.py) Performs ANOVA-based statistical analysis on EEG variability across stroke patients.

If you find this work useful, please cite our article:

Bunterngchit, C., Baniata, L.H., Baniata, M.H., ALDabbas, A., Khair, M.A., Chearanai, T. & Kang, S. (2025). GACL-Net: Hybrid Deep Learning Framework for Accurate Motor Imagery Classification in Stroke Rehabilitation. Computers, Materials & Continua. 83(1). 517-536. https://doi.org/10.32604/cmc.2025.060368

Publicly available datasets used in the article:

Dataset 1 from Liu et al.: https://www.nature.com/articles/s41597-023-02787-8

Dataset link: https://figshare.com/articles/dataset/EEG_datasets_of_stroke_patients/21679035/5

Dataset 2 from Tianyu Jia, Dataset link: https://figshare.com/articles/dataset/EEG_data_of_motor_imagery_for_stroke_patients/7636301