AHF - CIFRE - ASMA GASMI
Responsable LIMOS : AUGUSTO VincentDébut du projet : 15 juin 2019 - Fin du projet : 31 décembre 2022
Projet piloté par le LIMOS
Encadrement de thèse CIFRE Ateliers du Haut Forez : Asma GASMI -- Sleep analysis is a research field that has received a lot of interest lately due to its importance in the prevention of frailty. To analyze sleep in more efficient way, we use data that can be collected in a non-invasive way using contactless sensors. The first step of this research was related to the development of a new approach to analyze sleep architecture using data from low intrusive sensors. We focused on sleep phase detection, i.e. Wake, Non-Rapid Eye Movement (NREM) and Rapid Eye Movement (REM). The problem is considered as a supervised classification machine learning problem. Five machine learning methods were benchmarked using the same data. Using a neural network with short memory, we reached a 91.34% concordance with the gold standard for sleep stage detection, i.e. the polysomnography. We also proposed a new indicator to evaluate the sleep quality based on these results. Using such indicator, we proposed in a second step an approach to predict the frailty onset of elderly persons using sleep data over several days from a clinical trial. Using supervised machine learning approaches, we were able to predict frailty onset with a 92.4% precision. The last step of this work is related to the detection of the exact time of the drift in the sleep habits. Sleep anomaly detection is crucial to understand patient’s sleep habits and health state. We propose a new method based on auto encoders to detect anomalies in patient’s sleep habits. To do so we created a synthetic dataset of 2500 patients divided having 5 different types of sleep. We trained our algorithm to recognize such type of sleep. Finally, an auto-encoder method was used to detect the anomalies using a Long Short-Term Memory algorithm. We conclude this work with a real case study to illustrate the application of such tools for gerontologists.
Organismes partenaires :
Financeur : None