Machine learning the inverse problem of electro-cardiography mid year results

Documented are the early results in our research for machine learning the geometry of the inverse problem of electrocardiography. Our early research has focused on a concentric circle model to describe the system, where a smaller circle representing the heart is put inside a larger circle representing the body. Using a forward model that was provided to us by Mark Andrews, we generated heart surface potentials (HSP) and body surface potentials (BSP) to use as training and testing data. Early testing focused on whether our machine learning (ML) model could predict HSP’s from BSP’s. With this being quite a simple task for a regression model results were very good with under 4% RMSE on testing data. We then moved on to creating a model that could encode for the geometry of the heart. The goal was to have our model predict both the x and y position of the circular heart and the position of the peak of the gaussian on the heart. These results were also very promising with an average RMSE for the prediction of the x and y position of the heart being about 0.6 and the prediction of the angle of the gaussian peak being 0.01.