Bi-atrial synthetic populations with geometry-conditioned conduction velocity for atrial fibrillation simulations

Background Synthetic atria are a powerful tool complementing in-silico studies of atrial fibrillation (AF) through the generation of large datasets. In general, a statistical shape model (SSM) is used to generate realistic bi-atrial anatomies, with fibrosis maps assigned at random to the left atrium (LA) to define arrhythmogenic substrate [1, 2]; this decouples fibrosis from anatomy and overlooks their interplay during adverse atrial remodelling. Objective We previously introduced a LA statistical shape and appearance model (SSAM) that integrates conduction velocity (CV) maps as the appearance component [3], capturing realistic shape-electrophysiological relationships. Here, we expand our pipeline using a probabilistic (posterior) modelling approach [4], to enable the generation of simulation-ready bi-atrial populations equipped with data-driven, geometry-conditioned LA CV maps. Methods Twenty anatomically diverse surface meshes were sampled from a bi-atrial SSM [5]. To represent these cases within our SSAM, the LA endocardial surface was extracted, and represented as a set of correspondence points generated via a fixed-domain point-distribution modelling approach [6]. Our SSAM is trained to learn a joint latent space over geometry and CV measured during AF. We conditioned our SSAM on each new case to infer CV coefficients using the maximum a posteriori estimate as the predicted appearance field (CV). We truncated the latent space to the first 13 principal components (80% cumulative variance), to regularise inference and supress high-order noise. The predicted CV maps were then transferred back to the original bi-atrial meshes, preserving anatomical landmarks and inter-atrial connections required for activation modelling. To assess physiological plausibility of synthetic CV maps, we compared whole-LA median CVs and the percentage surface area with slow CV (<0.3 m/s) between the synthetic cohort (n=20) and our original SSAM training cohort (n=49) (Mann–Whitney U tests). Results Geometry-conditioned CV fields showed plausible regional heterogeneity visually, with localised slow-conduction zones (illustrative visualisations in Figure 1). Group medians of whole-LA CV were similar between cohorts (0.50 m/s [IQR 0.39 - 0.72] in the synthetic cohort vs 0.50 m/s [IQR 0.31 – 0.71] in the training cohort, p = 0.62; Figure 2A). Group medians of surface area with slow CV were also similar (32.2% [IQR 20.6 – 40.7] in the synthetic cohort vs 35.9% [IQR 21.5 – 48.6] in the training cohort, p = 0.35; Figure 2B). Conclusion Our enhanced SSAM pipeline enables the generation of bi-atrial, simulation-ready synthetic cohorts with realistic electrophysiological priors, improving hyperparameter tuning beyond random fibrosis maps. By reproducing realistic shape–CV relationships that result from coupled geometric and arrhythmogenic remodelling, it advances AF modelling and enables robust, large-scale in-silico studies of mechanisms and therapies.