Classification of Burst and Suppression in the Neonatal EEG

Abstract: The brain requires a continuous supply of oxygen and even a short period ofreduced oxygen supply risks severe and lifelong consequences for theaffected individual. The delivery is a vulnerable period for a baby who mayexperience for example hypoxia (lack of oxygen) that can damage the brain.Babies who experience problems are placed in an intensive care unit wheretheir vital signs are monitored, but there is no reliable way to monitor thebrain directly. Monitoring the brain would provide valuable informationabout the processes going on in it and could influence the treatment and helpto improve the quality of neonatal care. The scope of this project is todevelop methods that eventually can be put together to form a monitoringsystem for the brain that can function as decision-support for the physician incharge of treating the patient.The specific technical problem that is the topic of this thesis is detection ofburst and suppression in the electroencephalogram (EEG) signal. The thesisstarts with a brief description of the brain, with a focus on where the EEGoriginates, what types of activity can be found in this signal and what theymean. The data that have been available for the project are described,followed by the signal processing methods that have been used for preprocessing,and the feature functions that can be used for extracting certaintypes of characteristics from the data are defined. The next section describesclassification methodology and how it can be used for making decisionsbased on combinations of several features extracted from a signal. Theclassification methods Fisher’s Linear Discriminant, Neural Networks andSupport Vector Machines are described and are finally compared with respectto their ability to discriminate between burst and suppression. An experimentwith different combinations of features in the classification has also beencarried out. The results show similar results for the three methods but it canbe seen that the SVM is the best method with respect to handling multiplefeatures.