Atrial fibrillation is an increasingly prevalent condition in older people, but can be treated by radio frequency ablation. Computational models that describe the electrical behaviour of the left atrium in the heart can be used to predict outcomes for this procedure. The shape and function of the left atrium varies from one patient to the next, so calibration of these models to represent the left atrium of an individual patient is desirable. Imaging can be used to generate a mesh representing left atrial shape, and also to identify regions of fibrosis. However, measurements of electrical activation obtained in the clinical setting are typically sparse and noisy, and can be difficult to register to the mesh. Repolarisation cannot easily be measured during routine clinical studies, but effective refractory period can be estimated.
We have developed a workflow that uses probabilistic interpolation of measurements including local activation time and effective refractory period over the left atrial mesh. The probabilistic fields are then used to infer the distributions of model features (obtained by principal component analysis of model outputs) using a combination of approximate Bayesian computation and history matching. Our workflow has been tested using synthetic data, generated from simulations where the spatial variation in model parameters is known, and we have shown that both features and parameters can be recovered from simulated sparse measurements. The precision with which parameters can be recovered depends on the simulated pacing sequence, and so a further use of this workflow is to develop and assess novel pacing protocols.