Classification with a Deferral Option and Low-Trust Filtering for Automated Seizure Detection.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Wearable technology enables prolonged electroencephalography (EEG) monitoring for epilepsy patients at home.
• Manual visual analysis of 24-hour EEG recordings is time-consuming (1-2 hours) and prone to under-reporting of seizures by patients.
• Reliable automated seizure detection algorithms are essential to streamline EEG analysis and improve diagnostic accuracy.

Purpose of the Study:

• To develop an automated seizure detection methodology for behind-the-ear EEG data.
• To achieve high detection sensitivity (DS) by strategically deferring data segments to human experts.
• To enhance the performance of combined EEG segment prediction models by filtering untrustworthy data.

Main Methods:

• Investigated automated seizure detection algorithms using a dataset of behind-the-ear EEG measurements.
• Employed temperature scaling for prediction confidences and trust scores to determine data deferral.
• Implemented a strategy to filter untrustworthy EEG segments from combined predictions.

Main Results:

• A detection sensitivity (DS) of approximately 90% (99%) was achieved by deferring 10% (40%) of the data to human experts.
• Perfect DS was attainable when 50% of the data was deferred.
• Filtering untrustworthy segments reduced the false detection rate by 21-43% while maintaining or slightly improving DS.

Conclusions:

• Automated seizure detection algorithms, combined with expert review of challenging segments, can achieve high sensitivity for epilepsy diagnosis.
• Filtering low-trust segments in combined prediction models significantly reduces false positives, enhancing algorithm reliability.
• These advancements promise to reduce the burden of EEG analysis and improve the management of epilepsy.