Designing and Evaluating Mobile Health Technology for Ambulatory Monitoring and Diagnosis of Heart Arrhythmias

Cardiac arrhythmias comprise a large class of cardiovascular diseases (CVD). They are associated with an increased risk of heart attack and stroke and are a leading cause of global deaths. Atrial fibrillation (AF), a type of arrhythmia, alone affects over 1% of the worldwide population. It costs €73–95 million in the annual healthcare budget of Denmark and $26 billion in the USA. Electrocardiogram (ECG) analysis is a cost-effective and non-invasive means for detecting heart arrhythmias. However, due to their sporadic or paroxysmal nature, they often remain undetected in routine in-hospital ECG examinations, and therefore require longitudinal monitoring in patients under free-living conditions. Although longitudinal arrhythmia screening under free-living conditions can help in early diagnosis, it faces several challenges such as sustained patient engagement, poor signal quality, recall bias in the patient-reported symptoms diary or events, and a high false positive rate (FPR) in computer-aided automatic arrhythmia detection algorithms. Furthermore, in free-living ambulatory monitoring, motion artifacts often mimic arrhythmias and cause misdiagnosis. Without an understanding of the patient’s ambulatory context, it is difficult to ascertain if the ECG morphology is due to artifacts or arrhythmias. The high FPR in longitudinal screening increases the workload of clinicians (as it requires manual review) and could also lead to over-diagnosis and patient anxiety.
This dissertation investigates the role of context-awareness obtained via mobile and wearable devices to improve ambulatory arrhythmia analysis and reduce the FPR under free-living conditions. In addition, we also address the issues of recall bias in the patient-reported symptoms diary and events, and sustained patient engagement in longitudinal arrhythmia screening.
First, we identified the context information relevant for improving ambulatory arrhythmia screening under free-living conditions. After that, mCardia – a contextaware ECG collection system – was designed for longitudinal arrhythmia screening. We evaluated its usability and clinical feasibility in collecting contextualized ECGs for longitudinal arrhythmia screening in patients under free-living ambulatory conditions. Two clinical case studies from the collected contextualized data demonstrated the usefulness of contextual data in improving the manual analysis of ECG.
Furthermore, to improve the automated arrhythmia detection algorithm, we first investigated the influence of the patient’s ambulatory context on FPR in a state-of the-art arrhythmia detection algorithm. The investigation revealed that three specific ambulatory contexts, namely change in body position, activity change, and sudden movement acceleration, caused a significant number (62%) of non-trivial small segments of false positives. Based on these findings, we proposed a hybrid arrhythmia (AF) detection model named DeepAware. It employs deep learning and context-aware heuristics that significantly reduce the FPR under free-living conditions. When used in a clinical setting, DeepAware can significantly reduce cardiologists’ workload of manual review of FPs, allowing them to focus more on treatment than diagnostics.
With the two clinical case studies and performance of the DeepAware model, this PhD thesis demonstrated that contextual information could help in improving both manual and automated arrhythmia detection under free-living ambulatory conditions.
In addition, this dissertation also contributed a 259-day-long contextualized single-channel ECG arrhythmia dataset from patients under free-living ambulatory conditions. This database will help the broader deep learning community in building and evaluating the arrhythmia detection models that can realistically work under freeliving conditions. Furthermore, it will pave the way for making the deep learningbased, end-to-end arrhythmia detection models more explainable and will help identify the source of algorithm errors that otherwise remain a black box.