Design of a Multi-Feature Classification Scheme for Infant Epileptic Seizures.

Area of Science:

• Neurology
• Biomedical Engineering
• Signal Processing

Background:

• Neonatal epileptic seizures are a severe condition with high mortality and neurological sequelae.
• Accurate seizure detection in newborns is difficult, with a low percentage identified in neonatal intensive care units (NICU).
• Distinguishing seizures from non-cerebral activity is crucial for reliable diagnosis.

Purpose of the Study:

• To develop and evaluate a multi-feature approach for classifying neonatal epileptic seizures using EEG signals.
• To assess the performance of Support Vector Machine (SVM) and Random Forest classifiers for neonatal seizure detection.
• To identify significant spectral and statistical EEG features for improved seizure classification.

Main Methods:

• Analysis of EEG signals from 79 infants with suspected seizures.
• Extraction of spectral and statistical features from EEG data.
• Iterative training and assessment of SVM and Random Forest classification algorithms.
• Inclusion of an artefact reduction strategy to enhance signal accuracy.

Main Results:

• High classification performance achieved by both SVM (>80%) and Random Forest (>85%) models.
• Identification of nine high-scoring spectral and statistical features.
• Demonstration of the critical role of artefact reduction in improving classification accuracy.

Conclusions:

• A robust neonatal seizure classification scheme based on EEG spectral and statistical features is proposed.
• The identified features show potential as biomarkers for neonatal seizure prediction.
• Artefact reduction is essential for accurate automated seizure detection in neonates.