An AI-powered digital twin platform for predicting atrial fibrillation burden after catheter ablation

Background Atrial fibrillation (AF) burden is a promising biomarker for disease stratification and therapy guidance. Cardiac digital twins may offer a scalable approach to predict future AF burden pre-procedurally, improving patient selection and optimising ablation strategy. Purpose This work aimed to develop and validate a digital twin framework for predicting 12-month AF burden at the time of index catheter ablation. Methods and results Bi-atrial digital twins were constructed for 60 AF patients enrolled in the CASA-AF multi-centre randomized controlled trial. The trial randomized patients with long-standing persistent AF to catheter ablation (CA) or thoracoscopic surgical ablation (SA). AF burden was recorded over 12-months using implantable loop recorders. Late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) images were acquired before ablation and at 6 months follow-up. For each patient, two complementary digital twins were developed: a pre-ablation digital twin, derived from baseline LGE-CMR to predict future AF burden, and a post-ablation digital twin, derived from follow-up data to mechanistically interrogate clinical outcomes (Figure). First, the atrial anatomy was automatically segmented from the CMR scans using an ensemble of 3D residual encoder UNet models. Then, atlas-derived fibres were mapped onto the resulting atrial surfaces using Universal Atrial Coordinates. Native fibrosis and ablation scar were derived from the image intensity ratios (IIR) of baseline and follow-up LGE-CMR scans, respectively, and included in the corresponding digital twins. To predict 12-month AF burden at the time of index ablation, the planned ablation set was imposed on the pre-ablation digital twin. Conversely, the post-ablation digital twin incorporated the actual ablation scar (Figure: planned vs actual ablation scar). Burst stimulation was delivered from 10 clinically-relevant atrial sites and AF simulations were continued for 60 seconds. A vulnerability index was defined as the product of the number of induced AF episodes and their cumulative duration. The simulated vulnerability index in the pre-ablation digital twin closely matched the clinical AF burden, supporting its validity as a quantitative surrogate. Notably, discrepancies between the clinical outcomes and the vulnerability index derived from the pre-ablation digital twin were interrogated using the post-ablation digital twin. This comparison enabled identification of mechanistic drivers of recurrent AF, such as conduction gaps within ablation lines or residual fibrotic substrate not targeted by ablation. Conclusion This is the first computational framework using AI-powered Digital Twins to predict future AF burden in a large patient cohort and validated against 12-month clinical data. In addition to offering mechanistic insights into the factors driving AF recurrence, this framework can be leveraged to guide and optimise therapy.