Kernel machines for epilepsy diagnosis via EEG signal classification: a comparative study.

Area of Science:

• Computational neuroscience
• Machine learning in healthcare
• Biomedical signal processing

Background:

• Epilepsy diagnosis relies heavily on electroencephalogram (EEG) signal analysis.
• Automatic classification of EEG signals using machine learning can aid in diagnosis.
• Kernel-based learning machines offer powerful tools for complex pattern recognition in biological data.

Purpose of the Study:

• To systematically evaluate the performance of various kernel-based learning machines for epilepsy diagnosis.
• To compare the effectiveness of different kernel functions, parameter values, and feature extraction methods.
• To identify the most accurate and robust machine learning models for EEG-based epilepsy classification.

Main Methods:

• Investigated several kernel machines: standard Support Vector Machine (SVM), least squares SVM, Lagrangian SVM, smooth SVM, proximal SVM, and relevance vector machine.
• Conducted experiments on publicly available EEG data from normal subjects and epileptic patients.
• Evaluated performance based on predictive accuracy, sensitivity to kernel parameters, and feature types, using 26 kernel parameter values and 21 feature types derived from discrete wavelet transform and Lyapunov exponents.

Main Results:

• Assessed the impact of wavelet basis choice on feature quality, considering four wavelet functions.
• Reported average accuracy values for 252 kernel machine configurations, with standard and least squares SVMs achieving 100% accuracy in 40% and 35% of best-calibrated models, respectively.
• Demonstrated sensitivity profiles of configurations to feature types and kernel parameters, revealing critical decision-making factors.

Conclusions:

• All kernel machines demonstrated competitive accuracy, with standard and least squares SVMs being more consistently effective.
• The selection of kernel function, parameter value, and feature extractor critically impacts performance; wavelet family choice was less relevant.
• Statistical values from Lyapunov exponents were less informative than wavelet-derived features; optimal kernel parameter values and stable performance regions were identified.