Novel Approaches to Quantification and Segmentation of EEG Burst Suppression in Treatment of Status Epilepticus

Abstract

In this dissertation, novel possible approaches for burst suppression quantification and segmentation problem were proposed. The EEG burst suppression is a singular pattern that occurs in unconscious states and has been observed in coma with various etiology, general anesthesia, and the treatment of status epilepticus, which is analyzed herein. Burst suppression has been quantified using duration of bursts or suppressions and burst rates which are known to be associated to the depth of unconsciousness. Accordingly, correct, objective, and informative burst suppression quantification and segmentations has been studied. Previous studies have focused on developing and applying features to segment burst suppression accurately overcoming various problems in EEG. However, in this dissertation, three approaches were proposed considering diversification of segmentation process rather than a novel feature. Firstly, joint use of quantitative EEG features on time domain and frequency domain was proposed. Many features introduced for burst suppression segmentation or burst detection problems quantify signal consistency or inequality, and some of the features were also applied on power spectra in the frequency domain. The proposed method jointly uses the features applied to both EEG in time domain and power spectra in frequency domain. The generated two-dimensional features form a burst cluster and a suppression cluster in a two-dimensional domain. The clusters are modeled as Gaussian distributions and segmented by maximum likelihood estimation. Performance was evaluated by comparing the results with the visual segmentations of expert neurologists, and the proposed method produced a more accurate segmentation than conventional methods. Furthermore, applied pattern recognition technique was beneficial by reducing the subjectivity of many studies suggesting fixed thresholds. Secondly, an adaptive binarization for EEG burst suppression was proposed, and burst suppression segmentation conducted on the binarized EEG. One of the obstacles on EEG analysis is the wide dynamic range of EEG, which depends on the experimental settings and patients’ physical conditions. To resolve the problem, adaptive binarization on EEG and its power spectra was proposed restricting the wide dynamic ranges to 0 or 1, by borrowing constant-false-alarm-rate algorithm. For the binarized burst suppression, segmentation was performed by MLE with Gaussian modeling of feature-clusters. The proposed method showed higher agreement with the visual segmentation than the use of raw EEG. In addition, feature calculations on binarized EEG were simplified and saved processing time. Lastly, end-to-end segmentation and quantification through deep learning technology were proposed. Recent increasing technology, deep learning, is prevalent on various research field with its generalization power and end-to-end learning. Motivated by these virtues, long short-term memory, which is a sort of deep learning structure for time series analysis, was applied to burst suppression quantification and segmentation. Since the networks learned from visual segmentation of neurologists, the algorithm could not be compared to conventional methods directly, but the performance was found to be accurate. In addition, quantification of burst suppression without segmentation was also studied. Burst suppression ratio, one of the quantifications of depth of burst suppression, was estimated, and the estimation results roughly matched to target burst suppression ratio but required future work for fine estimation.