Deep Learning-Driven Identification of Atrial Fibrillation in the ICU
Loading...
Authors
Chen, Brian
Date
Type
thesis
Language
eng
Keyword
Deep Learning , Machine Learning , Precision Medicine , ICU , Atrial Fibrillation , ECG
Alternative Title
Abstract
The modern digitized intensive care unit (ICU) is an environment that relies on many sources of high frequency, continuously streaming sources of data for monitoring and decision making purposes.
One of the richest and most abundant sources is telemetry data from bedside monitors, and specifically electrocardiogram (ECG) waveform data.
Biosignals such as ECG hold a great deal of promise for precision medicine developments like predictive modelling or scientific discovery, especially of cardiopulmonary conditions.
One major condition of interest in the ICU is atrial fibrillation (AF) and in particular new-onset atrial fibrillation (NOAF).
AF is the most common arrhythmia among ICU patients and has been associated with a multitude of adverse outcomes.
Despite this, our understanding of the true burden of AF, how it manifests in an ICU context and how best to intervene in its treatment remain limited.
ECGs are an excellent way to identify AF, and ECG telemetry is very common in the ICU.
However, the majority of this ECG data is unlabelled, noisy and not as clean as the static, 12 lead diagnostic ECG that pervades most literature on machine learning with biosignals.
Although deep learning models have been shown to have state of the art performance in this domain, they require large quantities of labelled and more often than not clean data to effectively classify rhythms such as AF.
In this thesis, we describe three contributions to scale up AF detection in the absence of such data that cover the life cycle of a machine learning project.
Firstly, we create an annotator to speed up manual labelling with low effort to bootstrap the annotation process.
We then leverage deep learning models on noisily labelled data to classify AF, incorporating both noisy label generation methods and weakly supervised learning techniques to reduce the impact of noise.
Lastly, we look into quantifying the uncertainty of models trained and evaluated on our weak and human labelled ground truth data respectively to explore how such models might be interfaced with during deployment.
We conclude by discussing limitations of our approaches, as well as future work around proper clinical deployment for trials, real time monitoring and eventually prediction
Description
Citation
Publisher
License
Queen's University's Thesis/Dissertation Non-Exclusive License for Deposit to QSpace and Library and Archives Canada
ProQuest PhD and Master's Theses International Dissemination Agreement
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.
ProQuest PhD and Master's Theses International Dissemination Agreement
Intellectual Property Guidelines at Queen's University
Copying and Preserving Your Thesis
This publication is made available by the authority of the copyright owner solely for the purpose of private study and research and may not be copied or reproduced except as permitted by the copyright laws without written authority from the copyright owner.