Mapping populations to patients: designing optimal ablation therapy for atrial fibrillation through simulation and deep learning of digital twins

Atrial Fibrillation (AF) is an irregular heart rhythm that affects ~1 million people in the UK. It increases the risk of other cardiovascular diseases including heart failure, stroke and death. Patients who do not respond to drug therapy may be treated using radio frequency catheter ablation treatment, which aims to isolate areas of pathological tissue responsible for AF. In more advanced AF patients, treatment is sub-optimal with 40% of patients suffering from AF recurrence at 18-month follow-up. For these patients, 2-3 repeat procedures may be required. Different ablation approaches are used by different clinical centres and different clinicians, with varying degrees of personalisation to the individual patient anatomy, electrical properties and history.

In the current state of the art, there are large clinical population datasets available that inform treatment approaches for the average patient within a population. In parallel to this, there are also increasingly more detailed patient-specific models available. However, these approaches are typically disjoint. My vision is that we link information measured across a population to patient-specific models for predictive treatments. Tailoring ablation therapy to the individual patient, using data collected across a population, will improve long-term outcome and reduce recurrence, so that patients require fewer and shorter procedures.

Cardiac electrical signal mapping and imaging systems provide large quantities of spatial and temporal measurements for characterising the atria across populations of patients. These data can be used for constructing computational biophysical models of the atria, which in turn provide a physiological and physics constrained framework for investigating AF properties in personalised patient-specific models. Studying large virtual patient cohorts of these biophysical models can provide important insights into AF treatment approaches. However, these computer models run too slowly to be used during clinical procedures, and currently only capture what happens immediately after the treatment, and not the long-term response (for example, a year after the procedure). Machine learning techniques can capture complex relationships and generate fast predictions. To overcome the challenge of predicting the long-term response within a clinical timeframe, I will train a machine learning network to the large virtual patient cohort and clinical datasets to quickly predict treatment outcome from patient imaging and electrical data. I will also use imaging datasets that show how the structure of the atria changes during the months following the procedure. Updating the model (or digital twin) with these measurements will improve the ability of the model or network to predict the long-term outcome for the patient.

This project will move models from the research environment to clinical applications. These methodologies will be combined into a clinical tool that processes imaging and electrical measurements to map population therapy outcomes to patient-specific predictive treatments during the clinical procedure. The tool will take imaging and electrical data for a patient, and output different ablation therapy approaches together with how likely they are to reduce AF recurrence (and improve outcome), to aid patient-specific treatment planning. This project has clinical and industrial project partners to enable clinical translation of the technology. Use of the methodology developed during this fellowship may lead to better treatment selection, and decreased time and cost for AF catheter ablation procedures.