Recognize enhanced temporal-spatial-spectral features with a parallel multi-branch CNN and GRU

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalograms (MI-EEG) are complex, exhibiting non-stationarity and uneven distribution.
• Existing deep learning methods struggle to effectively fuse and enhance multidimensional MI-EEG features.
• Accurate MI-EEG recognition is crucial for advanced brain-computer interfaces.

Purpose of the Study:

• To develop a novel method for enhancing MI-EEG data representation and feature extraction.
• To improve the accuracy and reliability of motor imagery classification.
• To address the limitations of existing methods in handling complex MI-EEG characteristics.

Main Methods:

• A novel channel importance (NCI) approach based on time-frequency analysis was developed.
• The NCI method generates image sequences (NCI-ISG) by converting MI-EEG to time-frequency spectra, computing NCI, and creating weighted sub-band images.
• A parallel multi-branch convolutional neural network and gate recurrent unit (PMBCG) was designed for spatial-spectral and temporal feature extraction.

Main Results:

• The NCI-ISG + PMBCG method achieved average accuracies of 98.26% and 80.62% on two public four-class MI-EEG datasets.
• The approach demonstrated superior performance compared to state-of-the-art methods in MI-EEG classification.
• Statistical evaluations including Kappa value, confusion matrix, and ROC curve confirmed the method's effectiveness.

Conclusions:

• The proposed NCI-ISG method effectively enhances feature representation across time-frequency-space domains.
• The NCI-ISG + PMBCG framework significantly improves MI-EEG recognition accuracy, reliability, and discriminability.
• This study offers a promising advancement for brain-computer interface applications utilizing MI-EEG signals.