Master's Thesis : Machine learning techniques applied to sleep-disordered breathing diagnosis

Abstract

Sleep-disordered breathing (SDB) is a family of pathologies caused by a total or partial collapse of the upper airways. They are characterized by certain breathing patterns. One paradigm to identify these patterns is the one of the Nomics company that created the Jawac sensor, enabling to measure mandibular lowering. In 2008, in his PhD, Frédéric Senny developed for Nomics an automatic analysis based on this sensor to help doctors in their diagnosis. His algorithm proceeded in 2 phases, a first phase in charge of differentiating between wake and sleep from the Jawac sensor and a second phase finding Sleep-disordered breathing in these sleep zones. The first phase showed a number of limitations that had to be corrected over the years by a consequent post-processing by Nomics.

In recent years, deep learning methods have proven themselves in many fields such as image classification and music classification. This thesis is therefore intended to explore these methods in order to try to find one that will improve sleep/wake classification based on the Jawac.

In this thesis, we start by identifying the limitations of the current Nomics algorithm to motivate the new approach. Then, we build a pipeline to build supervised learning datasets related to the Jawac signal. Next, we test algorithms from Random Forest (RF), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Convolutional Recurrent Neural Networks (CRNN) on this problem in order to find the best architecture. We show that the best performances are obtained from a CRNN type network inspired from the literature with which we finally get 95.82% of AUC (Area Under the Curve) on the test set used. For comparison, Senny's algorithm only obtains 83.27% of AUC on the same test set.