Cardiac digital twins (CDTs) represent a significant advancement in precision medicine. Nevertheless, effective personalization remains a major challenge.
We propose a framework to personalize ventricular digital twins from clinical data by combining computed tomography (CT) and electrocardiogram (ECG) through optimization of the His insertions to Purkinje network. A 3D ventricular pediatric model was reconstructed, and candidate insertion sites were defined at the Purkinje network at the center of each American Heart Association (AHA) region. The personalization strategy consisted of two phases guided by the Dynamic Time Warping (DTW) metric. First, we perform spatial optimization (SO) to identify the most suitable insertion locations using a genetic algorithm. Second, we conducted temporal optimization (TO) to refine their activation times through Bayesian Optimization. To reduce initialization bias, five independent realizations using different random seeds were performed. Monodomain simulations employed the O'Hara action potential model, and the ECG was computed using an infinite volume conductor approach.
For analysis, we considered the top 10th percentile of spatial optimization solutions, ranked by a combined score of DTW and mean cross-correlation aligned to V1 (CC). Within this subset of 99 simulations, the methodology achieved a mean Pearson's correlation coefficient of 0.73 ± 0.045 with respect to the clinical ECG. Temporal optimization added little benefit, underscoring that spatial localization with synchronous stimulation is sufficient for clinical QRS reproduction. Our findings show that different spatial configurations can yield similar QRS morphologies, although inter-lead delays may persist or even worsen after temporal optimization.
In conclusion, this framework enables the construction of CDTs from standard clinical data and suggests that spatial distribution of His-Purkinje insertion has a stronger influence than temporal sequences in reproducing ECG morphology. Moreover, the same Purkinje network, when combined with different spatial arrangements of insertions, can reproduce equivalent ECGs, highlighting the inherent uncertainty in His-Purkinje network modeling.