Background: Machine Learning (ML) methods have seen an explosion in their development and application across a huge variety of problems. Although interest is growing, the impact of ML in the field of electrocardiographic imaging (ECGI) still remains limited. Published literature is scattered and there is no common ground description and or perspective to enable comparison of proposed methods.
Methods: We address this limitation by reviewing some common ECGI using the formalism of probabilistic graphical models (PGM) to provide a common ground to compare ML methods to well known traditional (“physics-based”) approaches, as well as to each other. We will start by describing the traditional methods from the point of view of statistical models that use maximum a posteriori estimation that are equivalent to Tikhonov and Total Variation regularization, formalizing them as single latent variable models. We will then add some recent modifications of these methods and show how they lead to similar models but with increased complexity. This discussion will include methods that use simple spatial priors, temporal models with propagation dynamics, models designed to identify specific markers of disease and models of the uncertainty of the forward model. Finally, we will discuss which approaches have traditionally been used to do inference on these models. The result will be a general picture of how ML methods can in principle be used to expand the range of tools available for both modeling and solving ECGI problems, including where both deep neural nets and other variational and sampling methods may become applicable as the methods in ML become more mature.
Conclusions: A review of ML approaches that can be applied to ECGI will help contextualize all the efforts from different groups. The review will help consolidate the literature and support future efforts to incorporate new developments in ML to research in ECGI.