Energy-Efficient Tree-Based EEG Artifact Detection.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electroencephalogram (EEG) artifacts can be misidentified as seizures, increasing false alarm rates in epilepsy monitoring.
• Developing robust artifact detection is essential for accurate and reliable seizure detection systems.

Purpose of the Study:

• To implement an energy-efficient EEG artifact detection algorithm on a parallel ultra-low-power (PULP) platform.
• To improve the accuracy and reduce false alarms in epilepsy monitoring systems.

Main Methods:

• Utilized the TUH EEG Artifact Corpus dataset, focusing on temporal electrodes.
• Employed an automated machine learning framework for optimal feature extraction in the frequency domain.
• Parallelized and optimized artifact detection algorithms for a PULP embedded platform.

Main Results:

• Achieved 93.95% accuracy and a 0.838 F1 score using 4 temporal EEG channels.
• Demonstrated a 5.21x improvement in energy efficiency compared to existing low-power artifact detection methods.
• The model, combined with seizure detection, enables 300 hours of continuous monitoring on a 300 mAh battery.

Conclusions:

• The developed EEG artifact detection framework significantly enhances the robustness of epilepsy monitoring.
• This technology facilitates affordable, wearable, long-term epilepsy monitoring with high sensitivity and low false-positive rates.
• The framework is suitable for integration into wearable EEG devices for improved seizure detection scenarios.