Learning Physics and Physiology Based Cardiac Activation Imaging

Aims of the Project:

Learn surrogate models for predicting body surface electrograms
Calibrate surrogate models using patient-specific anatomy and body surface electrograms
Learn lead locations for measuring activation in the whole or regions of the heart
Use scar images to improve activation imaging
Validate activation imaging
With each beat, the heart is activated starting in the top two chambers (atria) before activating the lower two chambers (ventricles). This coordinated activation regulates the efficient and effective pumping of blood by the heart. The breakdown of the activation pattern can cause severe cardiac dysfunction and understanding the timing and location of activation is critical in diagnosis and planning therapies.

The standard approach to measuring activation in the heart is an ECG but the interpretation of ECGs can be challenging, and ECG do not provide a detailed picture of activation patterns. Ground truth activation patterns can be measured invasively but this comes at a cost to the health system and risk to the patient. We can image the motion, shape, and tissue of the heart but these do not provide robust information for inferring activation patterns.

Current techniques for inferring ventricle activation rely on large numbers of electrodes to be placed on the torso to measure the body surface potential. By treating the body as a conductor this body surface can be used to estimate the potential on the surface of the heart. This problem is poorly posed, and so regularising methods are used to constrain the problem. The electrode vests are expensive limiting the wider adoption of activation imaging. The current ECG imaging (ECGi) approach ignores basic physiology, provides no estimate of uncertainty, and cannot be combined with other imaging data.

An alternate approach is to create a patient-specific physics-based model from all available data and calibrate this model to patient imaging and electrical recordings. King's College London is a leading centre for this type of modelling approach; however, these models are large and expensive to solve, making conventional simulation approaches incompatible with activation imaging.
In this project, we propose to exploit developments in machine learning modelling to learn optimal low-cost approximations for high-cost and fidelity computational models of electrical activation of the heart and ECG. By training model surrogates against detailed physics and physiological-based models, we will be able to both infer the activation and offer mechanistic explanations for the inferred activation pattern. The project will first develop a prototype electrical imaging workflow using simulated data. Then this approach will be applied to retrospective clinal data. Finally, we will prospectively compare inferred and invasively measured activation patterns in patients receiving a clinically indicated procedure.

This project will combine image and signal analysis, modelling and simulation, and machine learning and AI technique to create and test a novel cardiac activation imaging system.