A hybrid unsupervised and supervised learning approach for postictal generalized EEG suppression detection.

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Sudden unexpected death in epilepsy (SUDEP) is a major cause of mortality in epilepsy patients, often linked to uncontrolled seizures.
• Postictal generalized electroencephalogram (EEG) suppression (PGES) is a recognized risk marker for SUDEP, but its accurate detection is challenging due to physiological artifacts in real-world epilepsy monitoring unit (EMU) data.
• Existing automated PGES detection methods struggle with artifact-laden EEG recordings, hindering reliable SUDEP risk assessment.

Purpose of the Study:

• To develop and evaluate a novel hybrid machine learning approach for accurate PGES detection in multi-channel EEG recordings.
• To address the challenge of physiological artifacts in EMU data that impede precise PGES endpoint determination.
• To improve the reliability of SUDEP risk stratification through enhanced PGES detection.

Main Methods:

• A hybrid approach combining unsupervised (K-means clustering) and supervised (Random Forest) learning was developed for PGES detection.
• K-means clustering was used to group EEG recordings based on artifact features, enabling tailored model training.
• Random Forest models were trained using a novel strategy informed by clustering results to enhance detection performance.

Main Results:

• The hybrid approach achieved a 64.92% detection accuracy with a 5-second tolerance and 79.85% accuracy with a 10-second tolerance.
• The average predicted time distance for PGES events was 8.26 seconds.
• Leave-one-out cross-validation on 286 EEG recordings demonstrated superior performance compared to existing methods.

Conclusions:

• The proposed hybrid unsupervised and supervised learning method significantly improves PGES detection accuracy in challenging, artifact-prone EEG data.
• This approach offers a more robust tool for SUDEP risk assessment by providing more reliable PGES duration measurements.
• The findings suggest a promising direction for advancing automated analysis of EEG data in epilepsy monitoring and SUDEP research.