Predicting arrhythmia recurrence post-ablation in atrial fibrillation using explainable machine learning: Atrial meshes

This repository contains computational meshes of human left atria segmented from late gadolinium enhanced cardiac magnetic resonance imaging (LGE-MRI) for patients receiving a catheter ablation for atrial fibrillation. There are two models per patient (pre-ablation with patterns of fibrotic remodeling, post-ablation with scar created by the procedure) for 82 individuals. These meshes support the replication of our findings while protecting confidential patient information.

Methods
This study retrospectively included patients from University of Washington (UW) Medical Center with documented persistent atrial fibrillation (AFib) or paroxysmal AFib who had already received both pre- and post-procedural LGE-MRI scans and underwent either cryoballoon or radiofrequency (RF) ablation. Cardiac late gadolinium enhanced magnetic resonance images (LGE-MRI) were obtained using previously described protocols for all participants within 90 days prior to their ablation procedure and again 3-6 months post-ablation to quantify the extent of LA fibrosis and scar, respectively. Exclusion criteria for AFib patients included those who had a prior catheter ablation, patients with cardiac implantable electronic devices, severe claustrophobia, renal dysfunction, and contraindications to MRI or gadolinium-based contrast. Scans were performed on the Philips Ingenia system, 15–25 min after contrast injection, using a three-dimensional inversion-recovery, respiration-navigated, ECG-gated, gradient echo pulse sequence. Acquisition parameters included transverse imaging volume with a voxel size of 1.25 × 1.25 × 2.5 mm (reconstructed to 0.625 × 0.625 × 1.25 mm). Scan time was 5–10 minutes dependent on respiration and heart rate.

Patients had clinical assessment and catheter ablation in the UW AFib program. All patients underwent pulmonary vein isolation (PVI), and some had additional substrate modification at the operator’s discretion. 

Geometric models were reconstructed from both pre- and post- ablation LGE-MRI scans by Merisight Inc. (Salt Lake City, UT) to assess LA volume and surface area. Geometric models were reconstructed from pre-ablation scans and the relative extent of fibrosis in the LA was quantified via an adaptive histogram thresholding algorithm to determine pre-ablation LGE-MRI derived fibrosis. For post-ablation models, ablation scar was quantified on post-ablation LGE-MRI using previously established methods. Non-rigid registration was used to map LGE-derived post-ablation scar patterns onto existing LA pre-ablation fibrotic models. Hyper-enhancement on post-ablation scans was assumed to be ablation-induced scar; this accounts for the fact that hyperenhancement from ablation scar is at a higher absolute level than that of native fibrosis. Consequently, regions labeled as fibrotic pre-ablation fall below the hyperenhancement threshold in post-procedure scans.