A Multi-Source Transfer Joint Matching Method for Inter-Subject Motor Imagery Decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Individual differences present a significant challenge for accurate motor imagery (MI) decoding in brain-computer interfaces (BCI).
• Existing multi-source transfer learning (MSTL) methods often aggregate data from multiple subjects, potentially overlooking crucial sample information and inter-subject variability.

Purpose of the Study:

• To develop and validate advanced MSTL techniques, specifically multi-source transfer joint matching (MSTJM) and weighted MSTJM (wMSTJM), to mitigate individual differences in MI decoding.
• To introduce an inter-subject MI decoding framework that effectively utilizes these novel MSTL algorithms.

Main Methods:

• Proposed MSTJM and wMSTJM methods align data distributions for each subject pair, followed by decision fusion, unlike previous approaches that combine all source data.
• Developed an inter-subject MI decoding framework incorporating Riemannian covariance matrix alignment, Euclidean space source selection, and distribution alignment via MSTJM/wMSTJM.

Main Results:

• The MSTJM and wMSTJM methods demonstrated superior performance compared to state-of-the-art techniques on public MI datasets.
• Average classification accuracy improvements of at least 4.24% for MSTJM and 2.62% for wMSTJM were observed.

Conclusions:

• The proposed MSTJM and wMSTJM algorithms effectively reduce individual differences in MI decoding.
• The developed framework and MSTL methods show significant promise for advancing the practical applications of motor imagery-based brain-computer interfaces.