A deep neural network with subdomain adaptation for motor imagery brain-computer interface.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) signals exhibit significant nonstationarity, particularly spontaneous signals.
• This nonstationarity hinders the effectiveness of machine learning models in tasks involving spontaneous EEG, limiting practical applications in brain-computer interfaces (BCI).

Purpose of the Study:

• To develop a novel transfer learning algorithm to enhance classification accuracy for motor imagery (MI) EEG signals.
• To leverage previously labeled MI EEG data to improve performance when limited labeled data is available at the current time.

Main Methods:

• Introduced an adaptive layer within a deep convolutional neural network's fully connected layer.
• Designed an objective function to minimize Local Maximum Mean Discrepancy (LMMD), prediction error, within-class distance (DWC), and maximize between-class distance within each domain (DBCWD).
• Validated the algorithm on two public EEG datasets.

Main Results:

• The proposed algorithm achieved higher classification accuracy compared to existing methods.
• Paired t-tests confirmed statistically significant performance improvements.
• Confusion matrix and feature visualization corroborated the algorithm's effectiveness.

Conclusions:

• The transfer learning algorithm demonstrates superior performance in classifying MI EEG data with limited current labeled samples.
• The proposed method shows significant promise for practical applications in the field of BCI.