Improved neonatal seizure detection using adaptive learning.

Area of Science:

• Medical Technology
• Neuroscience
• Pediatrics

Background:

• Continuous electroencephalogram (EEG) monitoring is crucial in neonatal intensive care units (NICUs).
• There is a critical need for real-time EEG interpretation to identify neonatal seizures, particularly brief ones.
• Existing automated seizure detection algorithms struggle with short-duration neonatal seizures.

Purpose of the Study:

• To improve the performance of a multi-stage neonatal seizure detector.
• To enhance the detection of brief-duration seizures (< 30s) in neonates.
• To incorporate clinical neurophysiologist feedback for adaptive threshold tuning.

Main Methods:

• A three-stage neonatal seizure detector was developed, building upon a previous heuristic and data-driven model.
• A novel third stage was introduced to adaptively retune the second stage's threshold using clinical feedback.
• Performance was evaluated based on false alarm rate (FAR) and positive predictive value (PPV).

Main Results:

• The false alarm rate (FAR) for brief seizure detections decreased by 50%.
• The positive predictive value (PPV) for brief seizure detections increased by 18%.
• Overall FAR decreased by 35% and PPV increased by 5% with no change in good detection rate.

Conclusions:

• The addition of a third adaptive stage significantly improves the accuracy of neonatal seizure detection, especially for brief seizures.
• The refined algorithm offers a more reliable tool for around-the-clock EEG interpretation in NICUs.
• This advancement holds promise for better management of neurological conditions in neonates.