Subject-Independent EEG Classification of Motor Imagery Based on Dual-Branch Feature Fusion

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Brain-computer interface (BCI) systems enhance interaction for individuals with motor impairments.
• Current BCI systems face challenges in practicability and usability, often requiring extensive calibration.
• Calibration is time-consuming, depleting patient energy and causing anxiety.

Purpose of the Study:

• To propose a novel motion-assisted method for decoding human brain motion imagery intentions.
• To introduce a central loss function to improve classification accuracy by considering intra-class coupling.
• To achieve zero-calibration for motor imagery BCI (MI-BCI) systems.

Main Methods:

• Development of a dual-branch multiscale autoencoder network (MSAENet).
• Implementation of a central loss function to address limitations of traditional classifiers.
• Validation of the proposed method on three diverse datasets: BCIIV2a, SMR-BCI, and OpenBMI.

Main Results:

• The MSAENet demonstrated strong performance across all three datasets.
• In subject-independent scenarios, MSAENet outperformed four other methods on BCIIV2a and SMR-BCI.
• Achieved an F1-score of 69.34% on the OpenBMI dataset, showcasing effectiveness.

Conclusions:

• The proposed MSAENet offers improved classification accuracy with fewer parameters and shorter prediction times.
• The method successfully achieves zero-calibration for MI-BCI systems, enhancing practical application.
• This advancement significantly contributes to the usability and accessibility of BCI technology for motor dysfunction.