Deep Learning for Improving the Detection, Prognosis, and Treatment of Atrial Fibrillation

Abstract

Atrial fibrillation (AF) is the most common abnormal heart rhythm, contributing to stroke in one in five cases for people aged over 60 years and increasing the risk of heart failure three-fold. The majority of patients initially develop paroxysmal AF which is sustained for less than seven days. If left untreated, these patients can progress to persistent (and eventually permanent) AF, where the AF episodes last much longer and confer substantially higher risks of stroke. Persistent AF leads to progressive alterations in the electrical and structural properties of the atria which make restoring normal rhythm more difficult. Early detection and intervention are therefore needed to minimise the impact of AF and improve the quality of life for patients. However, the diagnosis of AF in clinics using a Holter monitor by trained physicians is difficult and error-prone, particularly due to its intermittent nature. Furthermore, current treatment performs poorly for persistent and longstanding-persistent AF, even when using the most promising ablation treatments which target localised atrial substrates via a minimally invasive procedure to treat AF. The low success rate of ablation in these groups is primarily due to the lack of effective clinical tools to detect and visualise the underlying atrial substrates sustaining AF. It follows that there is an urgent need for more intelligent methods to address these problems. Artificial intelligence is now widely applied in medicine. The superior data-driven capabilities of convolutional neural networks (CNNs) over traditional approaches have led to their becoming the dominant driver of artificial intelligence in the past decade. The overriding objective of this thesis was to develop novel CNN-based algorithms to automate and improve AF detection and treatment. Firstly, we addressed the lack of efficient methods for AF diagnosis by designing a convolutional recurrent network (CRN) for portable electrocardiogram (ECG) recording devices. The CRN was designed to process ECGs of varying input lengths by combining the feature learning capabilities of CNN and the recursive capacity of recurrent neural networks. A novel neural style transfer and mathematical modelling pipeline was designed to simulate synthetic ECGs and increase the training samples for the CRN. The framework was tested using the world’s largest open-source ECG dataset. We demonstrate that our approach detected AF with higher accuracy than commercial devices with significantly lower rates of false AF detection than current methods. Secondly, we addressed the lack of effective methods for detecting atrial substrates by developing a CNN-based pipeline for automatic structural analysis from 3D late gadolinium-enhanced magnetic resonance imaging (LGE-MRI). This was capable of performing segmentation and the extraction of key biomarkers including the bi-atrial chamber measurements, anatomic structure, fibrosis distribution, and wall thickness variation. The pipeline was trained and validated on the world’s largest LGE-MRI datasets. It was further tested on independent clinical data and demonstrated its efficacy by performing a range of analytical tasks with a higher degree of confidence and efficiency than traditional methods. Thirdly, we improved on current clinical mapping techniques for 3D atrial visualisation of the atria during ablation procedures. Present commercial tools combine 3D electrical mapping data with additional imaging and manual registration to achieve accurate anatomic reconstructions. We developed a novel CNN for automatic 3D atrial chamber reconstruction directly from sparse point clouds acquired during mapping. The approach was validated using clinical MRI and computed tomography datasets, and demonstrated accuracies equal to or better than commercial software. Importantly, the CNN was capable of predicting anatomical structures in sparsely mapped regions, allowing accurate reconstruction of 3D atrial anatomy from smaller datasets. This new approach potentially accelerates and improves 3D atrial visualisation needed for ablation and could therefore reduce the time needed for these procedures. Overall, the study presented significant contributions to improving the detection, prognosis, and treatment of AF by developing intelligent algorithms which can be integrated into the current clinical workflow. The methods and results derived from this research could facilitate clinical decision making and improve clinical outcomes for patients suffering from AF worldwide.