Obstructive sleep apnea detection using fourth level devices

11/06/2020

Obstructive sleep apnea is a common sleep disorder characterized by an interruption of one’s breathing during sleep. Due to the high cost, complexity, and accessibility issues related to polysomnography, which is the gold standard test for apnea detection, it is desirable that an automation of the diagnostic test be based on a simpler method. A fourth level device, which has only one or two channels allows the study to be performed outside the sleep laboratory with a low level of intrusion to the patient. The obstruction or reduction of airflow normally decreases the blood oxygen saturation level which can be used as a marker for apnea detection. An investigation of obstructive sleep apnea showed that apnea events have progressive bradycardia, followed by abrupt tachycardia on the resumption of breathing. Thus, heart rate variability could be a complement to the oxygen saturation signal. In this PhD work, two systematic reviews were performed to analyze the already existing algorithms to detect obstructive sleep apnea. One review focused on the use of different sensors while the other review focused on deep learning, in order to discover future trends and knowledge gaps in the field. Two different methods are followed in this work: handcrafted featurebased methods (using shallow networks) and automated feature-based methods (using deep networks). With regards to the handcrafted feature-based methods, it is necessary to find a subset of features that obtain the highest accuracy for the classifier. As for the automated feature-based methods, though the features are chosen automatically during the training, the structure of the deep neural network (namely, the choice of hyperparameters) plays a crucial part in the results. Therefore, for both methods, these criteria are investigated. In most of the cases, the proposed studies can achieve better results compared with those found in the literature. Using the classified apnea events, apnea patients (were classified global classification) which consequently also performed better than those in the literature. Among the different methods implemented in this work for handcrafted feature-based methods, a combination of certain classifiers performed better than single classifiers. In relation to automated feature-based Convolutional Neural Network (CNN) methods, the multi-objective method produced better results. To reduce the high simulation time, a greedy algorithm with a very similar performance was developed.

Authors in the community:

• 

  S

  Sheikh Shanawaz Mostafa

  ist427013

Supervisors of this institution:

• 

  J

  João Paulo Salgado Arriscado Costeira

  ist12390

101503946

Doutoramento em Engenharia Electrotécnica e de Computadores

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico