Digital twins of patient ventricular electrophysiology require subject-specific representation of the ventricular activation sequence to be mechanistically accurate and predictive. Simulation-based inference is emerging as a promising method to map activation properties from the 12-lead ECG and imaging-derived ventricular anatomy.
In this study, we evaluated the accuracy of a simulation-based method to infer biventricular activation times from the 12-lead ECG against 98 monodomain activation benchmarks. Each benchmark comprised a unique patient ventricular anatomy (32 controls, 66 hypertrophic cardiomyopathy) and heterogeneous activation properties including myocardial fibre architecture, from which the simulated QRS complexes were used as target ECGs. During inference, ventricular activation times were reconstructed from the ECG by iteratively refining early activation sites and conduction velocities in a fast-endocardial Eikonal model framework until model and target QRS complexes matched.
Inferred ventricular activation sequences showed strong agreement with ground truths (Spearman r=0.74±0.19), with inferred model QRS complexes closely matching the target QRS complexes (r=0.91±0.04). Inference accuracy was similar between control and hypertrophic cardiomyopathy anatomies (r=0.70±0.21 vs. r=0.76±0.17, respectively), showing applicability of the method to diseased cases with substantial myocardial hypertrophy. Notably, an isotropic Eikonal model with a fast endocardial layer was sufficient to recover activation patterns even in the presence of ground truths with fibre anisotropy.
Our results demonstrate that activation patterns can be accurately reconstructed by the proposed simulation-based inference method from QRS complexes and the ventricular anatomy, even in cases with hypertrophy, showing that ECG-based inference can support subject-specific ventricular digital twins across diverse ventricular anatomies and activation properties.