Classification of 1D and 2D EEG Signals for Seizure Detection in the Newborn Using Convolutional Neural Networks

Abstract

Newborns do not always show clinical symptoms during seizures unlike adults. Therefore, uncontrolled seizures lead to serious brain damage. Timely detection of seizures plays a vital role for newborn babies. In this study, a deep transfer learning approach was proposed for automatic seizure detection on the C4-P4 channel using electroencephalography (EEG) signals of newborns. EEG signals have been used in 1D and 2D dimensions to ensure performance, robust functionality, and a clinically acceptable level of detection accuracy. Pre-trained deep learning models Alexnet, ResNet, GoogleNet and VggNet were used in the study. Spectrograms were obtained by converting 1-dimensional signal data to 2-dimensional images, and then the classification was made on both 1D and 2D data set. In 1D classification, the highest performance was obtained from VggNet architecture with 91.67%, while 2D classification was obtained from AlexNet and ResNet architecture with 95.83%. The use of spectrograms has greatly improved the classification performance and made seizure detection and decision clinically more reliable in newborns.