Segmentation automatique des cris des nouveau-nés en vue du dépistage précoce des problèmes neurophysiologiques

Abstract

Several Studies have established the existence of a large number of information in an infant cry signal. Based on this assumption, many researches are devoted to the analysis of the cry signal in order to classify in one hand, the type of cry (birth cry, pain, hunger, discomfort, etc.) and in other hand the physical state of the newborn. This thesis describes the development and validation of an automated segmentation tool for the detection of vocal expiration and inspiration phases of newborn cries collected in a noisy hospital environment. This tool will be part of the preprocessing phase of newborn crying signals, prior to the automatic pathology classification system for newborns. As a first step, we have contributed to the establishment of a healthy and pathologic newborns’ cries database, intended to be public, accessible at all times for multiple research purposes related to the health of infants. The implementation of the cry database fulfilled expectations, including: secure data archiving, easy, and fast retrieval of information by means of adequate and effective interface and fast downloading to different locations for different uses. A corpus of 1939 cry signals were collected. 769 babies participated of which 372 are suffering from various diseases such as respiratory diseases, cardiac diseases, and neurological diseases. In a second time, we used supervised Learning methods, Gaussian Mixture Models and Hidden Markov Models, for the design of the automatic cry segmentation tool. Given the variability encountered in a real cry signals database, this tool is able to detect useful part of cries from other acoustic activities registered as the sounds of medical equipment, speech, noises at various levels and silence. Several signal processing and recognition tools have been investigated in this work in order to offer a fully automatic cries signals segmentation module robust towards noise and applicable in a real clinical environment and the most important, does not require a static threshold. In order to exploit the information available in the cries signals in different ways, we have applied and compared the most used signal decomposition techniques namely Fast Fourier Transform, Wavelet Packet Transform, and Empirical Mode Decomposition. We extracted different features to characterize and model separately and efficiently each type of vocal expiration and inspiration. The third area of focus, and in order to improve the results obtained from the supervised approaches by reducing boundary detection errors of useful segments, we integrated a post-processing stage to take full advantage of the time information of the signal. The full architecture realized is based on two consecutive modules. The first module uses cepstral features and traditional statistical approaches to give first results’ classification, and the second uses time and frequency features to correct errors and improve overall results. The various proposed approaches were tested on a different training and testing corpuses. The 10-fold cross validation technique is used to evaluate and verify the effectiveness of the proposed systems. The results of various tests show the robust performance of the proposed algorithms.

Date	21 Oct 2016
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Chakib Tadj (Supervisor)