Three-stage transfer learning for motor imagery EEG recognition.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) detection using electroencephalogram (EEG) is crucial for neural rehabilitation and drowsiness assessment.
• Brain-computer interface (BCI) advancements require accurate and efficient MI intention detection from EEG.
• Cross-subject heterogeneity and limited EEG data hinder current EEG-based MI decoding algorithms.

Purpose of the Study:

• To develop a novel three-stage transfer learning (TSTL) method for improved EEG-based motor imagery detection.
• To leverage optimal transport theory to address data distribution discrepancies in EEG.
• To enhance classification performance on unlabeled target domains using labeled source domain data.

Main Methods:

• Proposed a Three-Stage Transfer Learning (TSTL) method comprising Riemannian tangent space mapping (RTSM), source domain transformer (SDT), and optimal subspace mapping (OSM).
• RTSM minimizes marginal probability distribution drift by mapping Riemannian space to tangent space.
• SDT and OSM reduce joint and marginal distribution differences between source and target domains via optimal transport and subspace mapping.

Main Results:

• The TSTL method achieved average accuracies of 72.24% and 69.29% on two public BCI datasets.
• Demonstrated improved performance in EEG-based motor imagery detection compared to existing state-of-the-art algorithms.
• Validated the effectiveness of the proposed method in mitigating cross-subject heterogeneity and data scarcity issues.

Conclusions:

• The developed TSTL method significantly enhances the accuracy and efficiency of EEG-based motor imagery detection.
• Optimal transport theory provides a robust framework for domain adaptation in BCI applications.
• This approach offers a promising solution for real-world neural rehabilitation and drowsiness detection systems.