CNS Damage Classification in Newborn Infants by Neural Network Based Cry Analysis

The central nervous system (CNS) of the human body is the whole system of brain, spinal marrow and nerve cells throughout the body that correlates and regulates the internal reactions of the body and controls its adjustment to the environment. It controls muscles and processes sensory information originating from visual, aural, and other sensorial systems. Apart from that, it constitutes important human properties such as the ability to learn, control, think, feel, have self-awareness et cetera. The central nervous system not only directs the body on a conscious level but also on a subconscious level, making sure that we keep breathing, for example. Considering all these functions, it is easy to understand that malfunctions of the central nervous system can easily result in very severe complications, ranging from diminished control over movements and actions to abrupt cessation of life. One of the possible ways for the central nervous system to become damaged is by a shortage of oxygen. Normally, oxygen is extracted from the air by the lungs, which have a huge surface that is rich with capillaries. This surface allows for diffusion of oxygen from the air into the blood. After this, the heart pumps the oxygen-rich blood through the body, where the oxygen is used for various purposes, including the function of the central nervous system. Hypoxia (or hypoxaemia) is the general phenomenon in which the oxygen level in the blood is too low for the body to perform normally. In hospitals, the oxygen level is usually monitored because it is a critical parameter in surgical anesthesia. Apart from the measurement tools available, hypoxia can often be diagnosed visually, because the subject's skin tends to turn bluish. In the case of newborns, however, this diagnosis is often more difficult. It is possible for the neonate to experience episodes of hypoxia before and during birth, in which case the hypoxic episode may very well go by unnoticed. When this happens, damage may have been inflicted to the central nervous system. This damage is often hard to recognize, as it may not show until a much later stage in the life of the newborn. In this paper we explore a neural network approach for classifying infant cry in order to detect hypoxia related CNS damage. The data set consists of 35 recorded cases of infant cries episodes, called the FCU data set. Each cry episode is split into valid cry units resulting in a data set of 183 elements, called the ACU data set. Relevant numerical features are determined by the research on features in the area of infant cry analysis and speech recognition. After determining the relevant features an ensemble of Radial Basis Neural Networks was constructed for both data sets (ACU and FCU) using bootstrap aggregation. Testing the classification performance resulted in a performance of 85 ±7 % (99 % confidence level) on the ACU data set and 76 ±19 % (99 % confidence level) on the FCU data set. These classification rates should be approached with caution, however. The 78% is the lower bound of the confidence interval for the results of the classification is based on a small data set. It should therefore be stressed that the results from this investigation are not to be taken any other than 'preliminary'. Much more data and testing needs to be done. In conclusion, the results that were found are encouraging in the sense that a statistically significant relationship was found. The feature selection and classification methods developed in this investigation can be a sound basis for continuation. Further research should divert its attention to the (frequency) measurement and noise issues and to the data availability issues KEYWORDS: infant cry analysis, neural networks, hypoxia-related CNS damag