A method for atoms selection applied to screening for sleep disorders

ROMÁN ROLÓN; LEANDRO EZEQUIEL DI PERSIA; HUGO LEONARDO RUFINER; RUBEN SPIES

Buenos Aires

Congreso; 1st Pan-American Congress on Computational Mechanics - PANACM 2015; 2015

International Association for Computational Mechanics

The Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) is a sleep disorder which consists in repetitive events of partial or total airflow decrease during sleep. This pathology has a 4% prevalence in the population around the world and, without appropriate treatment, it increases with age. Actually the gold standard for detecting OSAHS is a polysomnography in a sleep laboratory, which consists in the simultaneous measurement of dierent physiological signals. In the last years several research studies have shown that the pulse oximetry is a very attractive option of screening for OSAHS, since changes in the dynamics of oxygen in the blood stream (SaO2) can be related with respiratory problems. In the last fteen years, many dierent signal processing techniques were used for building appropriate representations of a certain types of signals. One of these techniques is known as "sparse representation". The idea behind the method is to represent the involved signal using only a few coecients in a certain dictionary, previously constructed. In this work the SaO2 signal is used in order to predict the occurrence of Apnea-Hypopnea (AH) events. First a dictionary is learned by using a statistical method (NOCICA), then a greedy pursuit algorithm is used in order to obtain the activation coecients. A subset of the most discriminative coecients is then selected and used as input of a pattern recognition neural network in order to classify AH events. The problem for finding the optimal dictionary and activation coecients gives rise to an inverse problem with sparse constraints. A multilayer perceptron with different number of inputs and neurons in its hidden layer is then tested and the optimal conguration is derived.