Single-Source to Single-Target Cross-Subject Motor Imagery Classification Based on Multisubdomain Adaptation Network

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Cross-subject motor imagery (MI) classification using electroencephalography (EEG) faces challenges due to time-related data distribution shifts caused by device and subject variability.
• These shifts increase domain differences in single-source to single-target (STS) tasks, leading to reduced classification accuracy.

Purpose of the Study:

• To propose a novel multi-subdomain adaptation method (MSDAN) to mitigate time-related data distribution shifts in EEG-based MI classification.
• To enhance classification accuracy in STS MI tasks by addressing domain discrepancies.

Main Methods:

• MSDAN employs adaptation losses calculated from both class-related and time-related subdomains, distinguished by data and session labels.
• It measures distribution differences between source and target subdomains to guide adaptation.
• The method concurrently minimizes adaptation and classification losses within its loss function.

Main Results:

• The proposed MSDAN was applied to the BCI Competition III-IVa dataset for STS MI classification.
• Experimental results demonstrated that MSDAN outperforms traditional domain adaptation and deep learning methods.
• MSDAN effectively resolves time-related data distribution problems, achieving higher classification accuracy.

Conclusions:

• The novel multi-subdomain adaptation method (MSDAN) offers a robust solution for time-related data distribution shifts in EEG-based cross-subject MI classification.
• MSDAN significantly improves classification performance in brain-computer interface applications, particularly for STS tasks.
• This method holds considerable value for advancing the accuracy and reliability of BCI systems.