Multimodal data fusion for real-time detection of obstructive sleep apnea
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Sleep apnea is a respiratory disorder that is characterized by interruption in breathing during sleep. Obstructive sleep apnea (OSA) is the most common form of apnea affecting about 29.4 million adults in the United States. It has been linked with multiple health complications including daytime fatigue, cardiovascular dysfunction, pre-operative, and post-operative complications, etc. Severe form of apnea can even be lethal if not treated timely. Yet about 80 percent of the OSA patients remain undiagnosed. Overnight polysomnography (PSG) is the current gold-standard of OSA diagnosis that has various limitations including expense, set-up complexity, and lack of facilities for conducting an overnight PSG. As a result, an alternative of laboratory PSG is highly desired. This study proposes an artificial intelligence (AI) based OSA detection model that is capable of detecting apneic activity in real-time without the supervision of a sleep expert using an optimum number of sensors. Electrocardiogram (ECG) and blood-oxygen saturation signals (SpO2) were collected for apnea detection from an open-source database. Multiple AI models were developed to detect apnea from the individual signals and their fusion. In addition, this work adopts depthwise separable convolution to build a lightweight and low-parametric apnea detection model from the raw signals without any signal pre-processing. The proposed model involves fewer floating-point operations resulting in improved energy efficiency per prediction. Moreover, the model operates on an overlapping processing window of 12s enabling per-second detection of apnea. Furthermore, this work performs a qualitative post-hoc visual explanation technique on the model and provides insights into the reasoning behind its decisions.
Table of Contents
DOI
PubMed ID
Degree
Ph. D.