Gaussian mixture models for classification of neonatal seizures using EEG

Area of Science:

• Biomedical Engineering
• Clinical Neurology
• Signal Processing

Background:

• Neonatal seizures are critical neurological events requiring prompt detection.
• Existing seizure detection methods may lack real-time capabilities or accuracy.
• Accurate and timely diagnosis is crucial for effective neonatal care.

Purpose of the Study:

• To develop and evaluate a real-time system for detecting neonatal seizures.
• To assess the performance of a Gaussian mixture model classifier for this application.
• To analyze system parameters and identify sources of misclassification.

Main Methods:

• Implementation of a real-time neonatal seizure detection system.
• Utilization of a Gaussian mixture model classifier.
• Inclusion of feature transformation and classifier output postprocessing techniques.
• Evaluation on a database of 330 hours of recordings from 20 neonatal patients.

Main Results:

• Achieved a mean good detection rate of 79%.
• Reported a low false detection rate of 0.5 false detections per hour.
• Provided a detailed analysis of parameter choices influencing detector performance.
• Identified patterns contributing to false detections through thorough review.

Conclusions:

• The proposed Gaussian mixture model-based system offers effective real-time neonatal seizure detection.
• The system demonstrates a favorable balance between detection rate and false alarm frequency.
• Further analysis of misclassified events can refine future detection algorithms.