Machine learning for the prediction of infection and evaluation of maturation in premature infants combining cardiac and respiratory variability ; Apprentissage automatique pour la prédiction de l'infection et de la maturation chez le grand prématuré en associant les variabilités cardiaques et respi...
This dissertation was framed in the Digi-NewB project, which was founded by the European Union, and had as main goal to improve health care for neonates through the development of new monitoring systems. The project involved partners from four countries, and collected health records, physiological signals, and video and sound data from infants in six hospitals in France. The objective of our research was to propose decision support systems (DSSs) based on machine learning models for the early diagnosis of LOS and for the evaluation of maturation in preterm infants. From the data types acquired in the project, we limited our scope to physiological signals. We focused on heart rate variability (HRV), respiration rate variability (RRV), and bradycardia data. The main contributions of this work are: (i) an assessment of the positive impact in the performance of machine learning models for the detection of LOS of including visibility graph indexes for the characterization of HRV; (ii) a high performing recursive neural network model for early diagnosis of LOS in preterm infants based on HRV features; (iii) an ensemble machine learning model for the evaluation of maturation of the infants in terms of their functional maturational age, derived from HRV, RRV, or bradycardia features; (iv) the validation of this model on a population that included preterm infants with normal and abnormal maturation. The models presented in this work serve as proof of concept for non-invasive DSSs that can have a high performance in real-time. ; Cette thèse s'inscrit dans le cadre du projet européen Digi-NewB dont l'objectif principal était de développer un nouveau système de surveillance des grands prématurés. Le projet a impliqué des partenaires de quatre pays et a permis de collecter des données cliniques, des signaux physiologiques, des données vidéo et des pleurs de bébés dans six hôpitaux en France. L'objectif plus spécifique de ce travail était de proposer des systèmes d'aide à la décision (DSS) basés sur des modèles ...