Combining EEG Features and Convolutional Autoencoder for Neonatal Seizure Detection

Area of Science:

• Medical Informatics
• Neuroscience
• Machine Learning

Background:

• Neonatal epilepsy is a critical condition frequently encountered in neonatal intensive care units (NICUs).
• Current detection methods often rely on supervised learning, requiring extensive labeled electroencephalogram (EEG) data.
• The need for efficient and accurate seizure detection in neonates is paramount for timely intervention.

Purpose of the Study:

• To develop a semi-supervised hybrid architecture, Fd-CAE, for enhanced neonatal seizure detection.
• To leverage unsupervised learning via convolutional autoencoders (CAE) for optimizing EEG feature representation.
• To improve the classification performance of neonatal epilepsy detection using a novel hybrid approach.

Main Methods:

• Extraction of time-domain and entropy-domain features from neonatal EEG signals.
• Training a convolutional autoencoder (CAE) on unlabeled EEG features for unsupervised representation learning.
• Utilizing the pre-trained encoder for feature learning on labeled data to achieve seizure classification.

Main Results:

• The Fd-CAE model achieved high discriminative ability on a neonatal EEG dataset.
• Performance metrics included 92.34% accuracy, 93.61% precision, 98.74% recall, and 95.77% F1-score.
• Unsupervised learning with CAE significantly enhanced the characterization and classification of EEG signals.

Conclusions:

• The Fd-CAE method demonstrates the efficacy of combining unsupervised feature learning with supervised classification for neonatal seizure detection.
• The proposed approach offers a promising solution for improving diagnostic accuracy in neonatal intensive care settings.
• This hybrid model effectively optimizes EEG feature representation, leading to superior seizure detection performance.