Advanced Artificial Intelligence, Medical Imaging, and Biophysical Modelling to Inform Therapy for Atrial Fibrillation

Abstract

Atrial fibrillation (AF) is a very common cardiac arrhythmia worldwide, linked to increased risks of stroke, dementia, and heart failure. Despite being the only curative treatment, catheter ablation (CA) therapy yields poor outcomes for cases of persistent AF. Computational modelling, including the development of digital twins, offers valuable tools for guiding and informing clinical decisions. Furthermore, artificial intelligence (AI) has demonstrated its capability to analyse complex datasets and predict patient outcomes with accuracy comparable to that of clinicians. Medical imaging also provides extensive data on a patient’s condition, adding another layer of insight. Therefore, by integrating these three approaches, there is significant potential to enhance the treatment and outcomes for patients with persistent AF. This thesis proposes that the application of medical imaging, image-based computational modelling and AI can lead to the creation of innovative methodologies for improving the effectiveness of AF therapy. The first study of the thesis developed a deep learning (DL) model to automatically segment the left atrium (LA) and fibrotic tissue from late- gadolinium-enhanced (LGE) cardiac magnetic resonance (CMR) images, enabling more efficient preprocedural planning and patient stratification. Using a dataset of 60 LGE-CMR images, the model achieved state-of-the-art accuracy and highlighted how LGE-CMR image quality affects segmentation accuracy. The second study focused on AI safety, specifically investigating model transparency and interpretability. A DL model was developed to predict in-silico CA outcomes using 122 unfolded 2D LA models derived from LGE-CMR images and 199 synthetic 2D LA models. Grad-CAM, LIME and Occlusions interpretability techniques were tested, with Grad-CAM demonstrating the most interpretable saliency maps. This was evaluated using three quantitative interpretability metrics developed to assess whether these methods were focusing on physiologically relevant features. The third study investigated in-silico the role of fibrotic border zone (FBZ) in AF dynamics, to understand its contribution to the mechanisms of AF. Using nine 3D LA models derived from LGE-CMR images, the study explored how the presence of slow-conducting FBZ impacts phase singularity reentrant drivers (PS-RDs) sustaining AF. The results indicated that such an explicitly modelled FBZ stabilised PS-RDs by driving them deeper into dense fibrotic tissue areas with low conduction velocity (CV). The final study combined 3D LA simulations and explainable AI to identify factors influencing PS-RD localisation in AF, aiming to improve the mechanistic understanding of AF and to assist in developing more effective CA strategies for persistent AF. By evaluating six functional and three structural features from 41 LGE-CMR-derived models used for AF and sinus rhythm simulations, a random forest AI model classified the PS-RD regions. Shapley additive explanations revealed low CV as the most important factor for classifying PS-RD regions, providing insights into likely CA targets for patients with persistent AF. In summary, this thesis investigates the potential of combining Advanced AI, Medical Imaging, and Biophysical Modelling to Improve CA Therapy for Persistent AF. This includes LA segmentation from patient images, image-based LA model simulations, mechanistic insights into therapy targets in AF patients, and their prediction with interpretable AI. After further clinical validation, the developed methodologies and insights can offer a path toward more effective and personalised AF treatment strategies.

Date of Award	1 May 2025
Original language	English
Awarding Institution	King's College London
Supervisor	Oleg Aslanidi (Supervisor) & Andrew King (Supervisor)