Instance Transfer Subject-Dependent Strategy for Motor Imagery Signal Classification Using Deep Convolutional Neural...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) face challenges due to inter-subject variability in electroencephalogram (EEG) signals, hindering classifier generalization.
• Subject-dependent (SD) approaches improve personalization but are limited by data scarcity, especially for deep neural networks (DNNs).

Purpose of the Study:

• To propose an Instance Transfer Subject-Independent (ITSD) framework combined with a Convolutional Neural Network (CNN) to boost classification accuracy in motor imagery (MI) tasks.
• To address the limitations of data scarcity in personalized BCI classification models.

Main Methods:

• Developed an instance transfer learning method using the perceptive Hash algorithm to quantify EEG spectrogram similarity across subjects.
• Integrated this instance transfer learning with a CNN for decoding EEG signals.
• Compared the performance of Subject-Independent (SI)-CNN, SD-CNN, and the proposed ITSD-CNN training strategies.

Main Results:

• Instance transfer learning demonstrated positive transfer capabilities within a CNN classification model.
• The ITSD-CNN achieved an average classification accuracy of 94.7 ± 2.6%, showing significant improvement over a contrast model (p < 0.01).
• The ITSD-CNN framework outperformed state-of-the-art methods, achieving a mean kappa value of 0.664 on the BCI competition IV-2b dataset.

Conclusions:

• The proposed ITSD framework effectively enhances classification accuracy for motor imagery tasks in brain-computer interfaces.
• Instance transfer learning combined with CNNs offers a robust solution for overcoming inter-subject variability in EEG data.
• This approach represents a significant advancement in BCI technology, improving performance and generalization capabilities.