Image-Based Stratification and Therapy Selection for Atrial Fibrillation Patients Using Deep Learning

Aims of the Project:

Develop image-based deep learning models for stratification of atrial fibrillation (AF) patients.
Apply the models to identify AF patients who would not respond to catheter ablation therapy.
Apply the models to identify AF patients with high risks of stroke who require anticoagulation.
Validate predictions using available data on clinical outcomes and follow-up in patients.

Atrial fibrillation (AF), the most common arrhythmia, affects over 50 million people worldwide and accounts for a third of ischaemic strokes. Its diagnosis and management pose a substantial burden on healthcare systems, warranting efficient clinical approaches for stratifying AF patient therapy selection and stroke risks. However, even advanced AF therapies, such as catheter ablation (CA), are highly empirical and have poor long-term outcomes, with arrhythmia recurring in about half of the patients. Clinical approaches to stroke risk assessment are also empirical, based on patient characteristics (such as age, weight) and comorbidities, and are mostly effective for high-risk AF patients. This warrants the development of novel, reliable approaches to AF patient stratification that can account for anatomical and functional metrics from imaging and clinical exams.

Image-based assessments are increasingly used to move away from empirical diagnosis and therapy and improve patient outcomes. Thus, fibrotic tissue identified in the left atrium (LA) using MRI has been used as both a biomarker of AF severity and a target for CA therapy. Stroke risk scores may also be improved by multi-modal cardiac imaging that capture LA shape, motion or blood flow, since all these characteristics are linked with the likelihood of blood coagulation and thrombus formation. However, even advanced imaging systems provide limited information on key factors underlying AF sustenance and AF-related thrombogenesis, and the success of image-based stratification of AF patients remains suboptimal. Moreover, analysis of large volumes of patient imaging data requires substantial human and computational resources, which hinder their application in a clinical setting.

The application of deep learning (DL) can help overcome such limitations: (i) DL models can be trained using data from multiple sources, including patient imaging, image-based modelling, clinical records and ex-vivo experiments; (ii) once trained, DL models can provide a fast tool to support clinical decision-making based on available patient data. Therefore, this project will develop novel DL models trained on a combination of imaging, modelling and clinical data from AF patients. Once trained, the models will provide fast and flexible tools to identify 1) AF patients who are unlikely to respond to CA therapy and 2) AF patients with high thrombogenic risks who require anticoagulation. These predictions will be validated against clinical follow-up data available from the patients.