HFO Detection in Epilepsy: A Stacked Denoising Autoencoder and Sample Weight Adjusting Factors-Based Method

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computational Biology

Background:

• High-frequency oscillations (HFOs) detected via intracranial electroencephalography (iEEG) are crucial biomarkers for identifying epileptogenic zones.
• Accurate HFO detection is essential for pre-operative epilepsy assessment but is challenged by manual feature subjectivity and class imbalance.
• Existing methods struggle to achieve satisfactory performance due to these limitations.

Purpose of the Study:

• To develop a novel, accurate method for detecting HFOs in iEEG data.
• To overcome the limitations of manual feature extraction and class imbalance in HFO detection.
• To improve the pre-operative assessment of epilepsy by enhancing HFO localization.

Main Methods:

• An adjustable threshold of Hilbert envelopes was used to isolate events of interest (EoIs).
• A stacked denoising autoencoder (SDAE) was employed for automatic feature extraction in the time-frequency domain.
• An AdaBoost-based support vector machine ensemble classifier with sample weight adjusting factors was utilized to differentiate HFOs from EoIs, addressing class imbalance.

Main Results:

• The novel HFO detection method demonstrated superior performance compared to existing methods.
• The method achieved improved sensitivity and a reduced false discovery rate in clinical iEEG data from 20 epilepsy patients.
• Detected HFOs effectively aided in the localization of seizure onset zones.

Conclusions:

• The proposed SDAE and ensemble classifier method offers an accurate and robust approach for HFO detection in iEEG.
• This technique effectively addresses the challenges of subjective bias and class imbalance in HFO analysis.
• The method shows significant potential for improving pre-operative surgical planning in epilepsy patients by accurately localizing seizure onset zones.