Comparative multimodal calibration of patient-specific atrial fibrillation models: Impact of imaging and electrophysiology data on arrhythmogenic substrate identification.

Personalised computational modelling of atrial fibrillation (AF) integrates patient-specific imaging and electrophysiology to identify arrhythmogenic substrates and to support ablation planning in research settings. The influence of calibration modality and interpolation method on these models is not well established. Nine persistent AF (PersAF) patients undergoing first-time pulmonary vein isolation (PVI) were studied. Left atrial models were derived from late gadolinium enhancement magnetic resonance imaging (LGE-MRI) using atrialmtk and calibrated with LGE-MRI, electrogram-derived voltage, or conduction velocity (CV) data. EGM data were interpolated with radial basis function and Gaussian process manifold interpolation. AF sustainability, phase singularity (PS) distribution and cycle length (CL) were compared pre- and post-in silico PVI. Spatial concordance of PS regions with structural and functional metrics was assessed using Dice similarity coefficient. Across 234 simulations, AF was sustained in 94.9% pre-PVI, with no post-PVI differences (P ≥ 0.31). CV-calibrated models showed the highest PS counts, whereas LGE-MRI models the lowest (9.23 ± 5.78 vs. 3.11 ± 1.07). PVI reduced PS count (7.23 ± 4.32 to 4.08 ± 2.98; P < 0.001) and prolonged CL (e.g. 165.8 ± 19.4 to 184.2 ± 22.8 ms for voltage; P ≤ 0.045). PS hotspots overlapped more with low-voltage (0.474 ± 0.195) and low-CV (0.448 ± 0.126) zones than with high LGE regions (0.310 ± 0.057). Radial basis function and Gaussian process manifold interpolation showed moderate agreement in PS localisation, being higher for voltage calibration than CV (0.58 ± 0.08 vs. 0.53 ± 0.07; P = 0.017). Calibration modality and interpolation technique significantly influence AF dynamics and ablation target identification in patient-specific left atrial models, highlighting the impact of multimodal calibration. KEY POINTS: Patient-specific left atrial computational models calibrated with different clinical data modalities [late gadolinium enhancement magnetic resonance imaging (LGE-MRI), voltage, or conduction velocity] produce different atrial fibrillation dynamics and predicted arrhythmogenic substrates. Calibration using conduction velocity data resulted in the highest number of rotational activities and wavefront break-up, whereas LGE-MRI-based calibration produced the lowest, highlighting the influence of functional conduction heterogeneity. In silico pulmonary vein isolation significantly reduced phase singularity burden and prolonged cycle length across all calibration modalities. Spatial overlap of rotational activity regions is greater with low-voltage and slow-conduction zones than with high LGE-MRI intensity regions, indicating limited concordance between structural and functional substrates. The choice of calibration modality has a larger impact on simulated atrial fibrillation dynamics and predicted ablation targets than the choice of electrogram interpolation method; as no single modality captures all aspects of the arrhythmogenic substrate, comparative evaluation across modalities is necessary to understand their individual contributions and to inform future combined approaches for personalised ablation planning.