Peridevice leaks (PDLs) following left atrial appendage occlusion can negate the protective effect of the procedure and are a challenge in clinical cardiology. In this study, we investigate how the morphology of the anatomy relevant to the procedure interacts with the risk of PDLs using pre-operative CT images. We construct 3D models of the anatomy involved in the transcatheter procedure from a dataset of 125 patients who underwent left atrial appendage occlusion, integrating an anatomical centerline that simulates the catheter trajectory during device placement and that is complemented with morphological features. Given the complex and variable nature of this anatomy, we utilize Graph Attention Networks to analyze these models. This architecture enables us to encode morphological features into a graph structure, capturing the intricate spatial relationships and dependencies. We predict the likelihood of potential PDLs and identify key morphological descriptors contributing to these predictions through attention scores. This method provides a standardized representation of the anatomy involved in the procedure, offering insights into the anatomical factors influencing PDL risk and aiding in pre-operative planning.