Application of identity vectors for EEG classification.

Area of Science:

• Biometrics
• Signal Processing
• Machine Learning

Background:

• Developing optimal electroencephalography (EEG) subject verification algorithms remains a challenge.
• Lack of consistent benchmarks hinders understanding of classification improvements.
• Advancements often introduce new feature sets, classifiers, or datasets, complicating comparisons.

Purpose of the Study:

• To compare I-Vectors and Gaussian Mixture Model-Universal Background Models against a Mahalanobis classifier for EEG subject verification.
• To evaluate the impact of epoch duration on classification performance for different classifiers.
• To establish a consistent benchmark for EEG classification using a publicly available dataset.

Main Methods:

• I-Vectors and Gaussian Mixture Model-Universal Background Models were compared to a Mahalanobis classifier.
• Experiments utilized the publicly available PhysioNet database.
• Feature sets included spectral coherence, power spectral density, and cepstral coefficients.
• The effect of epoch duration on classifier performance was analyzed.

Main Results:

• I-Vectors demonstrated superior robustness compared to other classifiers.
• I-Vectors showed less sensitivity to variations in epoch duration, data composition, and feature selection.
• The I-Vector approach provided reliable baseline performance across different feature sets and datasets.

Conclusions:

• I-Vectors are well-suited for EEG signal processing tasks.
• This method helps standardize EEG classification by mitigating variations from feature sets and datasets.
• The I-Vector approach offers a reliable foundation for future EEG verification algorithm development.