Multi-Centroid Hyperdimensional Computing Approach for Epileptic Seizure Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Epilepsy patient monitoring requires reliable real-time seizure detection and prediction using wearable devices.
• Electroencephalogram (EEG) pattern variability poses significant challenges for traditional seizure detection algorithms.
• Existing methods struggle with high intra-class variability and imbalanced datasets common in real-world epilepsy monitoring.

Purpose of the Study:

• To develop a novel semi-supervised learning approach for improved epileptic seizure detection.
• To address the limitations of Hyperdimensional (HD) computing in handling high intra-class variability and data imbalance.
• To enhance the performance of epilepsy detection systems for long-term, continuous patient monitoring.

Main Methods:

• Proposed a novel semi-supervised learning approach utilizing multi-centroid Hyperdimensional (HD) computing.
• Developed multiple prototype vectors to represent both seizure (ictal) and non-seizure (inter-ictal) states.
• Evaluated the multi-centroid approach across three distinct dataset balancing scenarios, including highly imbalanced data.

Main Results:

• The multi-centroid HD computing approach demonstrated significantly improved performance compared to single-centroid models.
• Performance gains were most pronounced on highly imbalanced datasets, with up to 14% improvement observed.
• The method achieved better performance without a substantial increase in the total number of sub-classes.

Conclusions:

• The proposed multi-centroid HD computing approach offers a robust solution for epilepsy detection, particularly with imbalanced, real-world data.
• This method is a promising advancement for achieving high-performance seizure detection in continuous monitoring and online learning scenarios.
• The approach enhances the reliability of wearable devices for long-term epilepsy management.