EDSF-Net : An enhanced dynamic spatiotemporal-frequency attention network for robust EEG decoding in motor imagery

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) allows brain-computer interaction via imagined movements.
• Decoding MI from electroencephalography (EEG) is challenging due to neural complexity.
• Accurate MI decoding is crucial for advanced brain-computer interfaces (BCIs).

Purpose of the Study:

• To introduce an enhanced dynamic spatiotemporal -frequency attention convolutional neural network (EDSF-Net).
• To improve the precision and accuracy of motor imagery decoding from EEG signals.
• To address the challenges posed by neural pattern variability in MI detection.

Main Methods:

• Developed EDSF-Net incorporating enhanced dynamic convolution (EDConv) for spatiotemporal feature extraction.
• Employed a synchronized channel-frequency attention mechanism for focused feature learning.
• Utilized group convolutions formed by EDConv for effective feature fusion.
• Evaluated EDSF-Net on the BCI Competition IV 2a and OpenBMI datasets.

Main Results:

• Achieved 84.26% and 75.14% accuracy in hold-out session experiments on the two datasets.
• Attained 66.78% and 82.24% accuracy in leave-one-subject-out experiments.
• Demonstrated robust generalization capabilities and effective pattern recognition.

Conclusions:

• EDSF-Net significantly enhances motor imagery decoding accuracy.
• The model shows strong potential for diverse BCI applications.
• Advanced attention mechanisms and dynamic convolutions improve EEG signal analysis.