Systematic Assessment of Hyperdimensional Computing for Epileptic Seizure Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Hyperdimensional computing (HD computing) is a novel, low-power machine learning paradigm.
• HD computing shows promise for biomedical applications, including epileptic seizure detection.
• Inconsistent methodologies hinder comparative analysis of existing HD computing studies for seizure detection.

Purpose of the Study:

• To systematically assess the HD computing framework for epileptic seizure detection.
• To compare various feature extraction methods mapped to HD vectors.
• To evaluate generalizability across two distinct datasets.

Main Methods:

• Compared two existing and several novel feature approaches for HD computing-based seizure detection.
• Utilized a standardized preprocessing setup and identical performance metrics.
• Assessed detection performance, memory requirements, and computational complexity for wearable suitability.
• Evaluated a post-processing strategy to refine seizure prediction dynamics.

Main Results:

• Significant differences in detection performance were observed among feature approaches.
• High-performing methods often had substantial memory or computational demands, limiting wearable use.
• Post-processing substantially improved performance across all methods.
• Post-processing reduced performance disparities between different feature approaches.

Conclusions:

• The choice of feature extraction significantly impacts HD computing performance for seizure detection.
• Wearable implementation requires balancing detection accuracy with memory and computational constraints.
• A proposed post-processing strategy enhances the robustness and comparability of HD computing seizure detection systems.