Metric Learning in Freewill EEG Pre-Movement and Movement Intention Classification for Brain Machine Interfaces

Area of Science:

• Neuroscience and Biomedical Engineering
• Machine Learning and Signal Processing

Background:

• Decoding movement intentions from electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• EEG signals present challenges due to low spatial resolution and signal-to-noise ratio, hindering accurate decoding.
• Metric learning offers a method to learn data representations that capture desired similarities, potentially improving classification.

Purpose of the Study:

• To investigate the utility of metric learning in creating effective representations for classifying EEG movement and pre-movement intentions.
• To compare the classification performance of Support Vector Machines (SVM) using original (Euclidean) versus metric learning-derived data representations.
• To evaluate the interpretability of metric learning through a defined 'importance' measure for understanding data organization and channel contributions.

Main Methods:

• Employed three metric learning algorithms: Conditional Entropy Metric Learning (CEML), Neighborhood Component Analysis (NCA), and Entropy Gap Metric Learning (EGML).
• Utilized both time and frequency components as input features for the metric learning algorithms.
• Applied linear and non-linear Support Vector Machines (SVM) for classification accuracy comparison on a public EEG dataset (Subjects B and C).

Main Results:

• Metric learning algorithms did not significantly increase classification accuracies compared to the Euclidean representation.
• Metric learning provided interpretability via an 'importance' measure, revealing data organization and individual EEG channel contributions.
• Entropy Gap Metric Learning (EGML) demonstrated robust performance, effectively handling variable scale and correlations; feature selection (e.g., 0-5 Hz frequency range) significantly impacted results.

Conclusions:

• Metric learning offers significant benefits for understanding EEG data organization and feature importance in BCI applications, even without direct accuracy improvements.
• The 'importance' measure derived from metric learning provides visual explanations of data projections and inter-class separations, linking feature contributions to brain function.
• EGML shows promise for its robustness, and careful feature selection, particularly frequency range, is critical for optimizing metric learning outcomes and interpretability.