Estudo de algoritmos de aprendizado de máquina na predição de apneia do sono por sinais de EEG

Abstract

Sleep apnea is a sleep disorder characterized by partial or total obstruction of airflow in the respiratory tract. This leads to reduced blood oxygenation, inducing awakening. In the long term, it leads the person to develop other diseases that affect their health and social well-being. However, the diagnostic process is known and performed through the polysomnography, where bioelectrical and physiological signals are collected from the patient during a night’s sleep. Analysing bioelectrical signals is a complex task that involves signal processing; currently, however, there are computational tools in the field of artificial intelligence that aim to assist such analysis. In view of this, this research aimed to study a new technique for predicting sleep apnea, through the processing of only electroencephalogram (EEG) signals. To achieve this goal, EEG signals from the open access polysomnography database were used from the University Hospital St. Vincent, University of Dublin. This database is composed of polysomnography exam notes from 25 participants – 21 males and 4 females; mean age 50 (± 10) years, totalling 25 full-night polysomnograms (PSGs) with two EEG channels (C3-A2 and C4-A1). The database presents also notes of apnea respiratory events that occurred during the exam. The database was organized, grouping, in a single file, the EEG signal data and the annotations of respiratory events. With this data prepared, the signals were divided into frequency bands and the features that characterize them were extracted, in order to train the machine learning models. The study presented results that reveal the possibility of being able to identify sleep apnea only with EEG signals, thus facilitating the process of analysis and diagnosis of the syndrome. The Gradient Boosting algorithm, which had the best performance in this work, presented 79.56% of sensitivity, 90.59% of specificity and 84.96% of accuracy. Thinking about the continuity of the work, it is suggested to improve the generated models, focusing on the selection of attributes and the study of other algorithms. Finally, these studies can be a source of resources for practical applications, such as tools to support professionals in sleep apnea exams and studies.