Feature Selection for Automatic Burst Detection in Neonatal Electroencephalogram

Abstract

Monitoring neonatal electroencephalogram (EEG) signal is useful in identifying neonatal convulsions which might be clinically invisible. Presence of burst suppression pattern in neonate EEG is a clear indication of epilepsy. Visual identification of burst patterns from recorded continuous raw EEG data is time consuming. On the other hand, automatic burst detection techniques mentioned in the standard literature mostly rely on comparison with respect to predefined static voltage or energy thresholds, thus becoming too specific. Burst detection using ratio information of quantitative feature values between burst segment and neighborhood background EEG segment is proposed in this paper. Features like ratio of mean nonlinear energy, power spectral density, variance and absolute voltage, when applied as an input to a support vector machine (SVM) classifier, provides high degree of separability between burst and normal (nonburst) EEG segments. Exhaustive simulation using various literature specified features and proposed feature combinations shows that the proposed feature set provides best classification accuracy compared to other reported burst detection methods. The results documented in this paper can be used as a reference of optimum quantitative EEG feature sets for distinguishing between burst and normal (nonburst) EEG segments.