Dépistage de pathologies par analyse de cris néonataux à l'aide de réseaux de neurones

Abstract

This study focuses on the analysis of neonatal cry for the purpose of automated detection of pathologies. In this projet, we tried to accomplish this task using supervised neural network classification techniques. The recordings came from the École de technologie supérieure database. Only the cries of full-term infants aged 30 days or less were considered. The expiratory segments of the cries were separated into samples of 400 milliseconds, which were then subdivided further into 8 frames of 50 milliseconds. Each sample was then represented by the mel frequency cepstral coefficients (MFCC) of its frames. We evaluated three different neural network architectures : multilayer perceptrons (MLP), convolutional neural networks (CNN) and long short-term memory recurrent neural networks (LSTM). We trained these networks to recognize various pathologies, most notably hyperbilirubinemia and respiratory distress. k fold cross-validation was used to evaluate classifier performance. We also reproduced, as closely as possible, the dataset used in another study, in order to perform a fair comparison. We trained and evaluated our system on this dataset. The comparison of the performances we obtained with those reported by the reference study leads us to conclude that our approach has potential, but remains for the moment inferior to theirs. Finally, we also demonstrated that inadequate data partitioning between training and validation sets could produce a very large underestimation of the true generalization error. We raised suspicions about how the data were partitioned into several other studies on automated recognition of pathologies by analysis of newborn cries.

Date	12 Dec 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Chakib Tadj (Supervisor) & Christian Gargour (Co-supervisor)