Improved multi-stage neonatal seizure detection using a heuristic classifier and a data-driven post-processor.

Area of Science:

• Neuroscience
• Medical Technology

Background:

• Neonatal seizures require accurate detection for timely intervention.
• Existing heuristic classifiers for seizure detection have limitations in false alarm rates.

Purpose of the Study:

• To improve the performance of a heuristic seizure detection algorithm using a data-driven post-processor.
• To reduce false alarms while maintaining high seizure detection rates in neonatal electroencephalogram (EEG) data.

Main Methods:

• A heuristic classifier was used to identify seizure-relevant characteristics.
• A novel post-processor extracted features (synchronization, evolution, retention, segment, signal) from the heuristic output.
• A support vector machine and decision-making layer were employed to remove false positives.

Main Results:

• The heuristic method alone yielded a false alarm rate of 3.81/hour and 88% detection rate.
• The post-processor reduced the false alarm rate by 34%.
• The good detection rate decreased by a minimal 2%.

Conclusions:

• The data-driven post-processor significantly enhances the performance of heuristic neonatal seizure detection algorithms.
• The generic structure of the post-processor allows for application to other neonatal seizure detectors.
• This technique deepens the understanding of visually determined EEG features in neonatal seizures.