Effective and extensible feature extraction method using genetic algorithm-based frequency-domain feature search for...

Area of Science:

• Neuroscience and Biomedical Engineering
• Signal Processing and Machine Learning

Background:

• Epilepsy diagnosis relies heavily on electroencephalogram (EEG) signal analysis.
• Accurate feature extraction and selection are crucial for effective EEG-based epilepsy classification.
• Existing methods may not fully capture the complexities of EEG signals for optimal diagnostic performance.

Purpose of the Study:

• To propose a novel genetic algorithm-based frequency-domain feature search (GAFDS) method for EEG analysis.
• To enhance the classification accuracy of EEG signals for epilepsy detection.
• To evaluate the effectiveness of GAFDS in extracting independent and superior features compared to nonlinear features.

Main Methods:

• Developed the GAFDS method to search for frequency-domain features in EEG signals.
• Combined frequency-domain features with nonlinear features, followed by selection and optimization.
• Utilized Hilbert transformation to search for instantaneous frequency features.
• Employed multiple classical classifiers (k-NN, LDA, Decision Tree, AdaBoost, MLP, Naïve Bayes) to validate feature effectiveness.

Main Results:

• GAFDS extracted features demonstrated remarkable independence and superiority over nonlinear features (interclass vs. intraclass distance ratio).
• The method successfully identified instantaneous frequency features.
• Satisfactory classification accuracies were achieved across various classifiers, reaching up to 99% for 2-classification and 97% for 3-classification problems.
• Cross-validation experiments confirmed the effectiveness of GAFDS in extracting discriminative features for EEG classification.

Conclusions:

• The proposed GAFDS method is effective for feature extraction and optimization in EEG signal analysis for epilepsy.
• GAFDS significantly improves classification accuracy, demonstrating good extensibility and potential for clinical application.
• The feature selection and optimization model enhances the overall performance of EEG-based epilepsy diagnosis.