EEG motor imagery classification through a two-dimensional CNN-LSTM deep architecture and fuzzy decision-making

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) detection from electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Existing methods often struggle with noise and require complex feature engineering.
• Robust and accurate automatic detection remains a significant challenge.

Purpose of the Study:

• To develop a deep learning framework for automatic motor imagery detection from raw EEG signals.
• To enhance decision reliability in noisy EEG conditions through feature fusion.
• To improve the performance of motor imagery classification.

Main Methods:

• Extracted six band-power features using Short-Time Fourier Transform (STFT).
• Trained dedicated 2D Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models for each frequency band.
• Fused model outputs using a Choquet fuzzy integral for enhanced reliability.

Main Results:

• Alpha-band model achieved 88% accuracy; sigma-band model achieved 87.1% accuracy.
• The fused deep learning architecture reached 90.4% accuracy on the BCI IV-2a dataset.
• The system achieved 92.21% accuracy on the BCI IV-1 dataset, outperforming existing methods.

Conclusions:

• The proposed deep learning framework offers a robust and accurate method for motor imagery detection.
• The fusion strategy effectively enhances decision reliability in noisy EEG data.
• This approach significantly advances the state-of-the-art in motor imagery classification for BCIs.