Predicting Atrial Fibrillation Mechanisms Through Deep Learning

Atrial fibrillation (AF) is the most common cardiac arrhythmia, affecting over 1.1 million people in the UK alone, and is associated with increased risk of other cardiovascular disease, stroke and death. Patients who do not respond to drug treatment may be treated using radio frequency catheter ablation therapy, which is used to isolate the areas of pathological tissue responsible for AF. AF patients require different amounts of treatment: some patients require multiple procedures, with more extensive ablation strategies; while for others, a more simple isolation of the pulmonary veins using radio frequency catheter ablation is sufficient. Predicting whether an ablation treatment approach is a sufficient treatment for a particular patient is a clinical challenge, which if solved could improve safety of ablation procedures, and decrease time and cost for these procedures.

Computational biophysical simulations personalised to patient properties, including cardiac imaging and electrical data, may offer substantial insights into AF and how to treat it, but run too slowly to be used during clinical procedures. My objective is to develop a combined biophysical simulation and machine-learning network pipeline that accurately quantifies the likelihood of success of ablation therapies for an individual patient quickly enough for use during a clinical procedure, to guide ablation therapy. The machine-learning network will be trained to large quantities of biophysical simulated data to ensure that it correctly captures the physics and physiology of the system. The training will then be augmented with the complexity and reality of clinical data. Finally, the deep learning pipeline will be tested in a retrospective study.

We hope that this study will provide a proof of concept for this predictive pipeline. Our novel approach has the potential to revolutionise the field of predictive modelling for AF by constructing a pipeline that enables patient-specific treatment approaches to be developed and applied during a single ablation procedure. We hope that in the future, clinical and research centres will be able to use the trained machine learning network to predict the factors responsible for AF in an individual patient and the outcome of different ablation procedures. This may lead to improved safety of ablation procedures, better patient selection, as well as decreased time and cost for these procedures.

Persistent atrial fibrillation (AF) patients are a heterogeneous population: some patients require multiple procedures, with more extensive ablation strategies; while for others, isolation of the pulmonary veins using ablation (PVI) is sufficient. Identifying persistent AF patients where PVI will be a sufficient treatment remains a clinical challenge, which if solved could lead to improved safety, better patient selection, as well as decreased time and cost for procedures. Biophysical simulations personalised to cardiac imaging and electrical data may offer substantial insights into the mechanisms underlying AF, but run too slowly to be used during clinical procedures. My objective is to develop a combined biophysical simulation and deep learning network pipeline that accurately quantifies the likelihood of success of PVI for an individual patient quickly enough for use during a clinical procedure, to guide ablation therapy.

Methodology: We will simulate a virtual patient cohort covering the range of observed electrical and anatomical properties. These biophysical simulations will use the cardiac monodomain equation and the Courtemanche-Ramirez-Nattel atrial cell model, solved on meshes constructed from MRI images, with different fibrosis distributions, and repolarisation and conduction properties. The deep learning convolutional neural network will be trained to large quantities of post-processed data from biophysical simulations to ensure that the network captures the physics and physiology of the system. The training will then be augmented with the complexity and reality of clinical data. Finally, the deep learning pipeline will be tested in a retrospective study. We hope this study will provide insights into the mechanisms sustaining AF. We hope that different research and clinical centres will contribute to and make use of the trained network to predict patient-specific AF mechanisms and PVI ablation outcomes.

Atrial fibrillation (AF) is the most common cardiac arrhythmia, affecting over 1.1 million people in the UK alone, and is associated with increased risk of other cardiovascular disease, stroke and death. Patients often receive extensive ablation treatment and require multiple procedures.

This proposal aims to identify the subgroup of persistent AF patients who can be treated through pulmonary vein isolation alone, by using a novel combination of electrical recordings and medical imaging data, with biophysical simulation and deep learning techniques.

This research proposal has short and longer term impacts on the research community, industry, the general public, patients, clinicians, the NHS and the economy as follows:

Short term impacts:

1. Research community
This research proposal is relevant for basic science, engineering and clinical researchers, and has the potential to benefit and strengthen the UK electrophysiology research community. The virtual cohort of computational models that will be generated during this proposal will be made available to other researchers. These may be used by computer scientists to work on algorithm improvement for running simulations, as well as for testing other research hypotheses. The framework used for combining biophysical modelling data and deep learning techniques fits into a wider initiative to combine full physics simulations with machine learning techniques, and is a general approach that may have applicability to other diseases. We hope to show that simulation facilitated cross training can reduce data requirements, and provides an alternate means for developing neural networks. Discrepancies in predictions between model and clinically trained networks may highlight novel mechanisms that are not included in our current models that represent our understanding of the atria, which will in turn provide new hypotheses for testing experimentally.

2. Industry
Within the healthcare industry, there is a large international market and commercialisation potential for electroanatomic mapping software. As such, commercialising the pipeline that will be developed during this project has the potential to benefit economic competitiveness in the UK.

3. General public
This project may be of interest to the general public as an example of engineering in healthcare. Our research is of interest to students as it may encourage them to pursue a career in science or medicine.

Longer term impacts:
In the longer term, we hope that the outcomes of this research proposal will benefit:

4. Patients
The development of predictive software frameworks could improve patient treatment by indicating whether a patient is likely to benefit from a given procedure, reduce procedure times and improve safety by predicting whether a treatment is sufficient, and also suggest new treatment options following computational testing.

5. Clinicians
This research has the potential to improve understanding of the mechanisms underlying AF and so is of direct relevance to clinicians for understanding AF in an individual patient and how best to treat it. This increase in information and understanding could lead to improved risk stratification and more effective procedures. Our international collaborations mean we hope to affect healthcare in both the UK and overseas.

6. NHS
Reducing procedure times and decreasing the number of repeat AF procedures may lower mortality rate and hospital care costs, and reduce the financial burden on the NHS. To ensure the best possible healthcare in the UK, it is important for the UK to retain its competitive edge in the electrophysiology research field.

7. Economy
Improving outcomes for patients with AF will improve the productivity of AF sufferers as well as reducing the burden on their families and careers, which may lower unemployment and increase the economic competitiveness of the UK. This is especially important with an ageing population.