Kernel-Based Relevance Analysis with Enhanced Interpretability for Detection of Brain Activity Patterns.

Area of Science:

• Computational Neuroscience
• Biomedical Signal Processing
• Machine Learning in Healthcare

Background:

• Electroencephalography (EEG) is crucial for understanding brain activity.
• Identifying specific brain patterns from EEG data is challenging.
• Existing methods often lack interpretability or optimal accuracy.

Purpose of the Study:

• To introduce Enhanced Kernel-based Relevance Analysis (EKRA) for automatic EEG brain activity pattern identification.
• To enhance feature selection for improved accuracy and physiological interpretability.
• To provide a data-driven strategy leveraging joint information from neural responses.

Main Methods:

• EKRA utilizes two kernel functions for joint information analysis.
• A Centered Kernel Alignment functional optimizes parameter-free linear projection for discrimination.
• Two scenarios: feature selection for interpretation and enhanced feature selection for accuracy.

Main Results:

• EKRA successfully identifies relevant brain activity patterns from EEG.
• The method creates a relevant representation space, emphasizing salient features.
• EKRA outperforms state-of-the-art methods in motor imagery discrimination and epileptic seizure detection accuracy.

Conclusions:

• EKRA offers a novel, data-driven approach for EEG analysis.
• The method balances high discrimination accuracy with enhanced physiological interpretability.
• EKRA provides a valuable tool for advancing brain-computer interfaces and neurological diagnostics.