Self-supervised motor imagery EEG recognition model based on 1-D MTCNN-LSTM network.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Accurate motor imagery electroencephalography (MI-EEG) recognition is crucial for brain-computer interface (BCI) systems.
• Existing methods often require large labeled datasets, limiting their practical application.
• Developing models with high accuracy and generalization is essential.

Purpose of the Study:

• To propose a novel self-supervised MI-EEG recognition method.
• To reduce the reliance on extensive labeled training samples.
• To enhance the classification accuracy and generalization ability of MI-EEG recognition models.

Main Methods:

• A self-supervised learning approach using a one-dimensional multi-task convolutional neural network and long short-term memory (1-D MTCNN-LSTM) model.
• A two-stage process involving signal transform identification and pattern recognition.
• Transfer learning by fine-tuning the backbone network from the first stage to the second stage with minimal labeled data.

Main Results:

• The signal transform identification stage achieved over 95% accuracy, reaching up to 100%.
• The MI-EEG pattern recognition stage yielded accuracies of 82.04% and 87.14% with F1 scores of 0.7856 and 0.839 on BCIC-IV-2b and BCIC-IV-2a datasets, respectively.
• Demonstrated superior performance compared to existing methods.

Conclusions:

• The proposed self-supervised method significantly improves MI-EEG classification accuracy.
• This approach offers a promising solution for accurate MI-EEG classification in BCI systems.
• The method's effectiveness in reducing the need for labeled data makes it highly applicable.