Machine learning-based classification of physiological and pathological high-frequency oscillations recorded by...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• High-frequency oscillations (HFOs) are crucial biomarkers for identifying the epileptogenic zone (EZ) in epilepsy.
• Physiological HFOs in non-EZ regions can complicate accurate EZ localization.
• Developing methods to distinguish pathological from physiological HFOs is essential for precise epilepsy surgery planning.

Purpose of the Study:

• To develop and validate a machine learning classifier capable of distinguishing pathological HFOs from physiological HFOs.
• To evaluate the classifier's performance using features from multiple domains.
• To assess the impact of feature selection on classifier accuracy for HFO analysis.

Main Methods:

• HFOs were detected in focal epilepsy patients undergoing stereoelectroencephalography.
• 37 features across time, frequency, entropy, and nonlinear domains were extracted from each HFO.
• A fast correlation-based filter (FCBF) was used for feature selection, and a machine learning classifier was trained and tested on data from different hospitals.

Main Results:

• A large dataset of pathological and physiological HFOs was compiled from 26 patients.
• The classifier achieved high AUC values (0.95-0.98 in training, 0.82-0.90 in testing).
• The classifier using all extracted features outperformed the one using FCBF-selected features.

Conclusions:

• The developed machine learning classifier reliably differentiates pathological from physiological HFOs.
• This tool has the potential to enhance the accuracy of epileptogenic zone localization.
• The findings support the advancement of HFO analysis in clinical epilepsy management.