Atrial fibrillation (AF) is a complex cardiac disease impact-ing an ever-growing population and increases 6-fold the risk of thrombusformation. However, image based bio-markers to predict thrombosis inpresence of AF are not well known. This lack of knowledge comes fromthe difficulty to analyse and compare the shape of the Left Atrium (LA)as well as the insufficiency of data that limits the complexity of modelswe can use. Conducting data analysis in cardiology exacerbates the smalldataset problem because the heart cycle renders impossible to compareimages taken at systole and diastole time. To address these issues, wefirst propose a graph representation of the LA, to focus on the impactof pulmonary veins (PV) and LA Appendage (LAA) positions, givinga simple object easy to analyse. Secondly, we propose a meta-learningframework for heterogeneous datasets based on the consistent represen-tation of each dataset in a common latent space. We show that sucha model is analogous to a meta-classifier, where each dataset is charac-terised by specific projection in a common latent space, while sharing thesame separating boundary. We apply this model to the graph represen-tation of the LA and interpret the model to give novel time-dependantbio-markers related to PV and LAA configurations for the prediction ofthrombosis.