Unsupervised heterogeneous domain adaptation for EEG classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) classification faces limitations due to scarce training data.
• Existing domain adaptation methods often overlook data heterogeneity caused by diverse EEG devices.
• There is a need for methods that leverage auxiliary heterogeneous data for improved EEG classification.

Purpose of the Study:

• To propose a novel unsupervised heterogeneous domain adaptation method for EEG classification.
• To develop a model that effectively utilizes knowledge from auxiliary heterogeneous EEG data.
• To address the practical challenge of EEG classification where source and target data reside in different spaces.

Main Methods:

• Introduced the informative representation fusion (IRF) model for unsupervised heterogeneous domain adaptation.
• Learned data representations from both independent identically distributed (iid) and non-iid perspectives.
• Utilized hypergraphs for non-iid data, multi-layer perceptrons for iid data, and an attention mechanism for fusion.
• Employed maximum mean discrepancy to align source and target domain distributions using fused features.

Main Results:

• The proposed IRF model demonstrated effectiveness on multiple real-world EEG datasets.
• Experimental results confirmed the model's ability to handle heterogeneous and unsupervised cross-domain classification.
• The model achieved high classification accuracy in challenging, practical EEG scenarios.

Conclusions:

• The IRF model successfully addresses unsupervised heterogeneous domain adaptation for EEG classification.
• This approach is valuable for real-world applications where EEG data originates from different devices and settings.
• The study contributes to the advancement of EEG-based Brain-Computer Interfaces (BCIs) by improving classification performance.