Abstract:

Atrial shape has been linked to atrial fibrillation (AF), yet many aspects remain poorly understood. Recent evidence suggests that the right atrium (RA) affects AF vulnerability in both chambers, highlighting the need for bi-atrial analysis. However, open bi-atrial cohorts are scarce. Statistical shape models (SSMs) address this gap, but model variability only linearly and require surface correspondence. Recently, deep learning (DL)-based generative models have emerged as an alternative. We present a DL-based bi-atrial generative model, extending an existing signed distance function based auto-decoder framework for the left atrium (LA) to a dual-head network with shared trunk, trained on 344 patients from a public CT dataset. We investigated latent dimensionality effects on reconstruction and generation performance. The model produced anatomically realistic geometries with separated atria, variable shapes, and diverse LA appendage morphologies. Volumes and volume ratios aligned with clinically observed ranges. Reconstruction and generation improved with latent dimension, with dimension 64 best balancing complexity and quality where mean Chamfer distances were 1.2 mm for reconstruction and 2.5 mm (LA) and 2.2 mm (RA) for generation. LA reconstruction distances exceeded those in the RA, likely due to greater topological complexity and the comparatively lower RA data fidelity. Compared to the existing bi-atrial SSM, the DL model required higher dimension, possibly confounded by dataset differences. Overall, loss plateaued within 400 epochs for all models, suggesting early stopping for future work. In conclusion, the model can support virtual cohort generation for AF in silico studies; allowing conditioning on clinical metadata in future work.

Abstract:

While atrial fibrillation (AF) is currently the most common arrhythmia worldwide, finding theoptimal treatment for each patient remains a challenge. One option for treating AF patientsis pulmonary vein isolation (PVI). PVI is an established rhythm control strategy. However,PVI often fails for some patients, leading to high recurrence rates, especially for those withpersistent AF. To treat these patients in whom PVI failed, ablation outside the pulmonaryveins can be performed. However, finding the right spots for ablation is demanding and time-consuming. To better identify suitable regions for ablation, electrophysiological simulationscan be used to assess AF vulnerability before the ablation procedure. These simulationscurrently require substantial computational power, making them unfeasible for smallerinstitutions or hospitals. Therefore, this thesis uses machine learning (ML) to assess andaccelerate the vulnerability assessment of AF patient-specific models.The ML pipeline consists of a feature-based approach where global clinical featuresof the atrium were calculated based on late gadolinium enhancement magnetic resonanceimaging (LGE-MRI). In addition, local features around single stimulation points werecalculated. Data from 22 patients with AF were used to create personalized patient models.AF vulnerability of these 22 patients was then calculated using the PEERP pacing protocolwith the electrophysiological simulator openCARP.A random forest (RF) classifier was used with the previously calculated global and localfeatures to predict the results from the pacing protocol for inducibility. This algorithm wasfurther improved using hyperparameter tuning and different optimization strategies. Thetuned algorithm achieved a mean area under the curve (AUC) receiver operating charac-teristic (ROC) of 0.75 ± 0.03 in predicting point inducibility with 10-fold cross-validation.Discarding 50% of the test samples based on RF certainty analysis achieved a false negativerate of 5%. Using this classifier configuration, up to 10% of all stimulation points areexcluded, resulting in a 10% reduction in simulation time.To better assess the influence of different features on the RF classifier, we used SHAPto get an insight into feature importance. The analysis showed that especially local fibrosisdensity around stimulation points had the highest impact on the prediction (SHAP valuesfrom −0.15 to 0.1 ). Additionally, SHAP showed that the inducibility for a stimulation pointdecreases if it is moved to the atrial appendages.This work not only provides an approach to saving 10% of simulation time for assessingatrial fibrillation vulnerability but also uses SHAP to explain and assess predictions of theunderlying random forest classifier.