Correlation analysis between EEG parameters to enhance the performance of intelligent predictive models for the neonatal newborn sick effects

Abstract

Electroencephalogram (EEG) is a signal that measures the electrical activity of the brain. It contains some specific patterns that predict neuro-developmental impairments of a premature newborn. Extracting these patterns from a set of EEG records provides a dataset to be used in machine learning in order to implement an intelligent classification system that predicts prognosis of the baby. In a previous work, we proved that inter-burst intervals (IBI) found in the EEG records as well as low amplitude Burst predicts abnormal outcomes of the premature. According to this hypothesis, we defined 20 parameters in EEG signal at birth to propose an efficient automatic classification system that predicts a risk on cerebral maturation at birth that can lead to a pathological state at 2 years. In this paper, we use correlation analysis between the 20 EEG parameters to find the redundant sets of them and eliminate those that are less correlated with the class, thereby reduce their number. To do this, we calculate the correlation coefficients between all the attributes to find their correlation matrix. Next, we choose the attribute sets with a correlation greater than 90% to find the parameters that give close results. Then among these parameters, we find the correlation between each of them with the class to determine which is the less important to eliminate it. Finally, we reduce the number of parameters to 17, and enhance the accuracy of the proposed classification system from 88,4% to 93,2%. This system has a good sensitivity to predict the neurological status of preterm infants and can be used as a decision aid in clinical treatment.