A novel dual-step transfer framework based on domain selection and feature alignment for motor imagery decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) utilizing motor imagery (MI) require lengthy calibration.
• Transfer learning (TL) shows promise in reducing MI-BCI calibration time.
• Subject data distribution variability hinders TL effectiveness in MI-BCI.

Purpose of the Study:

• To develop an efficient transfer learning framework for MI-BCI.
• To address challenges posed by differing data distributions across subjects.
• To reduce calibration time and improve classification accuracy in MI-BCI.

Main Methods:

• Proposed a dual-step transfer framework integrating data alignment, source domain selection, and feature alignment.
• Employed a pre-calibration strategy (PS) for initial source-target domain alignment.
• Utilized a sequential reverse selection method with a dual model strategy for optimal source domain matching.
• Incorporated filter bank regularization common space pattern (FBRCSP) for enhanced feature extraction.
• Applied manifold embedded distribution alignment (MEDA) to refine support vector machine (SVM) predictions.

Main Results:

• The proposed framework achieved higher average classification accuracy than baseline methods on public and private datasets (84.12%, 79.91%, 78.45%).
• Demonstrated significant reduction in computational cost, approximately halved compared to the baseline.
• Effectively mitigated the impact of varying data distributions across subjects.

Conclusions:

• The novel dual-step transfer framework enhances MI-BCI performance by optimizing source domain selection and feature alignment.
• This approach effectively reduces calibration time and computational load while improving classification accuracy.
• The study highlights the potential of advanced TL techniques for practical MI-BCI applications.