Optimization Strategies Comparison for ECG-Based Ventricular His-Purkinje Root Node Inference
Please login to view abstract download link
Ventricular activation inference from surface ECGs is a key challenge in non-invasive cardiac digital twining. A common approach is to optimize spatial locations and activation times of His–Purkinje root nodes to reproduce a patient’s ECG [1]. The choice of the optimization method and the parameter space strongly affect fitting accuracy, prediction consistency, and computational cost. Here we present a first systematic comparison of prediction performances of different optimization strategies for calibrating cardiac EP ventricular activation. We compare 3 global optimization methods: tree-structured Parzen estimator (TPE), genetic algorithm (GA), and particle swarm optimization (PS) and one local finite-difference (FD) method. Optimizers were evaluated both as standalone approaches for joint spatiotemporal inference and in hybrid configurations combining global spatial search with local spatiotemporal refinement. All methods were assessed on a biventricular mesh with eikonal model with template-based transmembrane voltage recovery, combined with a lead field approach to recover bidomain ECGs [1-2] and compared using mean RMSE, Pearson's correlation coefficient (CC), and number of simulations to best fit across 5 inference runs. GA and PSO outperformed TPE in fitness and stability (Table 1). GA applied to just spatial optimization achieved high CC with a very low number of simulations, highlighting its efficiency when search space dimensionality was reduced. Optimizer choice and problem parametrization influenced inference speed and solution quality. The most efficient strategy was GA for spatial optimization followed by FD refinement, whereas PSO standalone achieved the best fit. This work could provide guidance for selecting optimization pipelines in cardiac digital twin studies and in-silico clinical trials.
