A Comparative Study on EEG Features for Neonatal Seizure Detection

Abstract

Epileptic seizure is one of the common neurological disorders, and its clinical manifestation is different from that of the adult as the neonate’s brain is not yet fully developed. In clinical practice, manual observation of EEG recordings to diagnose and identify epileptic seizures is expensive and time-consuming. Computer-aided diagnosis (CAD) tools will enable clinicians to examine the EEG expeditiously and effectively. In this study, we investigate 37 statistical and timefrequency domain features to discriminate between seizure and non-seizure EEG segments. The analysis is performed on the publicly available Helsinki University database. The significant features were identified by using the Wilcoxon rank sum test and then ranked using manual threshold area under the curve (AUC) value and eXtreme Gradient Boosting (XGBoost) feature importance method. The performance of the features was analyzed using XGBoost and support vector machine (SVM) classifier with fourfold cross-validation. We found that entropy plays a significant role in the discrimination of seizure and non-seizure segments. We achieved an average AUC of 0.84 and 0.76 using XGBoost and SVM classifiers, respectively. This study presents the significance of each extracted feature and will be beneficial to the neurologists for the continuous monitoring and diagnosis of seizures in neonates. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2022.