Automatic Classification Framework for Neonatal Seizure Using Wavelet Scattering Transform and Nearest Component Analysis

Abstract

Introduction

Neonatal seizure is a common neurologic disorder in neonates. The diagnosis of a neonatal seizure can be made clinically or with an EEG. However, the clinical diagnosis of neonatal seizures is difficult, particularly in critically ill infants, because of the multitude of epileptic and nonepileptic clinical manifestations. On the other hand neonatal seizure can be effectively detected using EEG recordings. Hence, there is a need for an electroencephalograph (EEG) based automatic diagnosis framework for neonatal seizure.

Methods

This work proposed a wavelet scattering transform (WST) and histogram-based nearest component analysis (HBNCA) based framework for classifying seizures and non-seizure neonate's EEG signals. The WST converts EEG signals into its translation invariant and deformation stable representation. The HBNCA method is deployed to find the effective wavelet scattering coefficients (WSC) for classifying seizures and non-seizures EEG signals. Then, various classifiers are used to identify the effectiveness of the features.

Results

The proposed framework is managed to get an average accuracy of 98.59% and 97.83% for a 1-second duration of EEG signal for repeated random subsampling validation (RRSV) and leave one out cross-validation (LOOCV), respectively.

Conclusions

The results are compared with the other state of art methods. The accurate classification from the 1-second duration of the EEG signal shows the potential of the proposed framework for reliable neonatal seizure classification.