Approximately 90% of strokes related to atrial fibrillation (AF) can be attributed to thrombus originated in the left atrial appendage (LAA). This cardiac substructure is highly heterogeneous among different individuals. The morphology of LAA can be qualitatively classified into different categories (chicken wing, cauliflower, windsock, cactus) based on simple morphological measures such as the number of lobes, surface irregularity and bending of the main lobe, extracted from medical images. There is controversy on the relation between this LAA classification and the risk of thrombus formation, ischemic stroke and cerebrovascular accidents. In addition, very simple haemodynamics measurements from noisy 2D echocardiographic images are used to characterize LA/LAA haemodynamics. During the past years, we have developed several computational tools to describe LAA morphology and haemodynamics with advanced image processing tools and computational blood flow simulations. Nevertheless, it is not straightforward to find patterns among the large quantity of multi-dimensional parameters and their relationship with clinical outcomes. Advanced statistical studies have been performed on the available data, resulting in interesting new knowledge relating some morphological parameters (e.g. ostium dimensions) with the presence of thrombus. The main goal of this project will be to apply machine-learning tools to find patterns including in silico haemodynamics parameters, morphological descriptors and clinical data such as the presence of thrombus. In addition, a ML-based framework should be developed to automatically select the optimal device settings for a specific LAA geometry. It will be in collaboration with clinicians at Hospital Clínic de Barcelona, Hospital Sant Pau de Barcelona and international partners such as the DTU in Denmark.
Supervisors: Oscar Camara, Jordi Mill, Xabier Morales