Classifying neonatal EEG

Other conference contribution, 2007

In spite of considerable medical progress during the last decades, the perinatal period is still one of the high-risk periods in any individual’s lifetime. In neonatal intensive care of today there is a serious lack of methods that allow continuous monitoring of cerebral function. While we consider it mandatory that good quality hospital care shall include facilities for continuous monitoring of respiratory and cardiac functions in severely ill patients we lack the same possibility when it comes to this most important organ of the body, the brain. The electroencephalogram (EEG) can provide information regarding the state of the brain, but is in its current form not suited for continuous monitoring. Not all neonatal EEG characteristics have been fully investigated or are fully understood, and the people with the necessary competence for interpreting them is not available at neonatal intensive care wards. Our approach is to design a decision support system suitable for continuous monitoring that uses classification algorithms to classify the EEG, for example as normal continuous, normal periodic and pathologic periodic. The EEG is a highly complex signal, and rather than estimating a single parameter, the focus has been on applying classification methods on ensembles of parameters that describe the characteristics of the EEG signal. These parameters have been chosen to enhance different aspects of the EEG signal, and by training classification algorithms with manually segmented data characteristic differences between these complex signals can be found. So far, various classification algorithms have been tested on the task of classifying segments of burst-suppression EEG (pathological periodic) into burst and suppression with rather satisfying results. As a next step we have planned to look into the classification of EEG signals as continuous and periodic, and classification of periodic signals as pathological or normal.