Article Preview
Top1. Introduction
A premature baby is physiologically vulnerable to have neuro-developmental problems according to his gestational age. The electroencephalogram (EEG) is the most often used criterion in predicting this risk. Electroencephalography (EEG) is a neurological diagnosis that measures the electrical activity of the brain via electrodes (usually 10-20 system) placed on the scalp. They are distributed symmetrically on both hemispheres with respect to the frontal regions, occipital and temporal (Wikström, 2011). In the newborn case, 9 electrodes are enough (Koolen et al., 2013). The place of the electrode can cause an artifact in the EEG recordings such as traces caused by the activity of the heart that can be detected in ECG (electrocardiogram) or traces caused by the activity of the eye that can be detected in EOG (electro-oculogram) (Mizrahi, 2007). An EEG is requested when there are abnormalities in neurological status such as brain dysfunction. It determines the presence of brain injury as well as neonatal seizures. For example, an EEG should be done to diagnose unexplained apnea, flushing of the face as well as the neurological status lethargic or hyperactive (Lombroso, 1985; Wikström, 2011). Normal neonatal EEG determines normal brain development and progress. Abnormalities can be detected by specific patterns such as continuity, waveform and wake-sleep cycle. There are gradual variations from discontinuity to continuity (Mizrahi, 2007). Continuity appears in wakefulness and discontinuity appears in sleep. There are many waveform definitions such as Beta, delta theta and alpha that can be considered normal according to the gestational age of the newborn. Awake EEG and active sleep shows a continuous trace consisting of mixed frequencies.
The first step of our work is to extract from the EEG signal features that are related to the prognosis of the premature newborn. We used an application called EEGDiag (Chauvet and Nguyen, 2013), which segments the EEG signal in phases of burst and inter-burst intervals (IBI) and detects in each phase features useful in predicting neuro-developmental risk. These features are presented by 14 parameters. We used a data set of 397 premature newborns, collected in Angers hospital, having each his EEG at birth and his diagnosis two years later, which is classified into: normal, sick or risky (risky means that it was not clear yet if the newborn will be sick or not). Detecting these 14 parameters in the 397 EEG recordings, we obtained a dataset that contains 397 feature vectors, each one has 14 inputs and one output which is normal, sick or risky. Then, we executed machine learning using one of the most popular intelligent models (multiple linear regression, linear discriminant analysis, artificial neural network and decision tree) to generate the best intelligent classification system that determines the prognosis (normal, sick or risky) of a premature new born according to the 14 inputs detected in his EEG signal at birth (Wu et al., 2008). In this work, we used 80% of the dataset for training and 20% for test to specify the performance of the found system by comparing its result with that given by doctors.
In the next section of this paper, we define EEG features that affect the prognosis of the newborn. In section 3, we present the dataset used in machine learning. In section 4, we execute machine learning by applying the 4 intelligent models: linear discriminant analysis, multiple linear regressions, artificial neural network and decision tree and we present their results. In section 5, we discuss a complete comparative study between the 4 intelligent models, to evaluate the performance of each one. In the final section, we present the conclusion and perspective of this work.