Revealing False Positive Features in Epileptic EEG Identification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• Epileptic seizure detection from electroencephalogram (EEG) signals is crucial for diagnosis and treatment.
• Current EEG classification methods often extract numerous features, leading to high dimensionality and potential loss of relevant information.
• Existing feature selection strategies primarily focus on classifier performance, sometimes overlooking the intrinsic relationship between features and EEG activity.

Purpose of the Study:

• To propose a machine learning-based pipeline that enhances the association between selected features and epileptic EEG tasks.
• To introduce a novel feature selection algorithm utilizing a knockoff filter to identify optimal feature subsets.
• To improve classification accuracy for discriminating between normal, interictal, and ictal EEG states.

Main Methods:

• Extraction of diverse temporal, spectral, and spatial features from raw EEG signals.
• Application of a novel knockoff filter-based feature selection algorithm to derive an optimal feature subgroup.
• Evaluation of the selected features using three distinct classifiers: k-nearest neighbor (KNN), random forest (RF), and support vector machine (SVM).

Main Results:

• The proposed method achieved high accuracy, reaching 99.93% for normal vs. interictal EEG discrimination and 98.95% for interictal vs. ictal EEG classification on the Bonn dataset.
• Demonstrated superior performance compared to state-of-the-art techniques.
• Achieved average sensitivity of 95.67%, specificity of 98.83%, and accuracy of 98.89% on the Freiburg dataset.

Conclusions:

• The developed knockoff filter-based feature selection pipeline effectively enhances the relationship between features and epileptic EEG data.
• The proposed approach offers a promising and accurate method for epileptic seizure detection and classification.
• The findings suggest significant improvements in EEG-based epilepsy diagnosis using advanced machine learning techniques.