Population-level emulation enables fast Bayesian calibration of cardiac digital twins

A population-level Gaussian process emulator calibrated cardiac digital twins for 26 left bundle branch block patients in 59.7–63.0 s per patient, versus 15 h with the full simulator.

• Why it matters: Patient-specific Bayesian calibration is too slow for routine workflows because it needs thousands of simulations per patient, and existing emulators are usually trained one patient at a time.
• What they did: The team trained one emulator using a 12-dimensional latent encoding of patient clinical features, then ran Markov chain Monte Carlo to calibrate seven closed-loop heart and circulation parameters to ten biomarkers, comparing emulator outputs to the full simulator.
• The result: Leave-one-out calibration matched simulator accuracy (median leave-one-out normalized RMSE 11.0% vs 10.7%; p = 0.745), with error increasing with distance from the cohort centroid (Spearman rho = +0.406) and 95% predictive intervals covering 78.5% of targets.

AI summary of the abstract · may contain errors · not medical advice · read the paper.