Abstract:
Neonatal seizures are one of the most frequent neurological events in newborn infants which
reflects a variety of pre or postnatal disorders of central nervous systems and usually indicates
serious neurological dysfunction. These seizures may have nonexistent or subtle clinical
manifestations, patterns like oscillatory or spike train start with very low amplitude and gradually
increase over time. This makes neonatal seizure detection very difficult and inaccurate if it solely
relies upon clinical observation. Although, it has been shown that the most accurate method for
their detection or diagnosis is visual interpretation of continuous multi-channel neonatal
Electroencephalogram (EEG) along with video by an expert clinical neurophysiologist, such
interpretation is extremely labor intensive, time-consuming, and importantly relies on special
expertise which is not available continuously around hospital neonatal intensive care units
(NICUs).
A reliable and accurate automated neonatal seizure detection and classification using multi channel EEG can be a very important supportive tool, particularly for the NICUs. However,
identifying a core set of features is one of the most challenges in the development of an automated
neonatal seizure detection. In most of the published studies describing features and seizure
classifiers, the features were hand-engineered (feature selected manually), which may not be
optimal and the results claimed from previously proposed automation techniques are less accurate
and unreliable. Furthermore, the system that can detect neonatal seizure and identification of the
seizure grade from neonatal EEG dataset has not been previously done.
In this thesis, the detection and grade identification of neonatal seizure from multi-channel EEG
signal was proposed using deep convolutional neural network models. The proposed system was
developed using MATLAB software. The multi-channel neonatal EEG became preprocessed,
segmented and the two dimensional matrix changed to raw waveform image with defined size prior
to feeding to the custom CNN and pre-trained Alexnet models. The developed system was capable
of detecting the neonatal seizure using binary classification as well as grade level identification
using multiple classification techniques. The test result showed that the Alexnet perform better
result during binary classification with accuracy of 92.6% and custom CNN performs better result
on grade level identification (multi-classification) with accuracy of 88.6%.
