From P to S: what can we learn from the 12-lead ECG for cardiac digital twins

The 12-lead ECG is recorded roughly 300 million times per year worldwide, yet most of its spatiotemporal information is discarded. This talk presents three efforts that combine the ECG with biophysical models to build patient-specific cardiac digital twins from non-invasive data.
First, we reconstruct the ventricular Purkinje network from the QRS complex as a probabilistic inverse problem solved with Bayesian optimization. The framework yields uncertainty estimates on network parameters and propagates them through downstream predictions, such as conduction-system pacing in patients with left bundle branch block and intraventricular conduction delay.
Second, we recover atrial activation from the P-wave. Building on a rapid anisotropic eikonal lead-field formulation, we introduce Δ-PoIssoNN, a physics-informed neural network that recasts the anisotropic eikonal equation as a Poisson problem, constraining solutions to satisfy the propagation model. It outperforms a standard Δ-PINN on synthetic data and produces physiologically consistent activations on patient P-waves.
Third, we address cardiac arrhythmias by reconstructing phase maps from a complex eikonal formulation, with phase observations obtained from the Fourier transform at the dominant frequency. A physics-informed neural network recovers re-entry and spiral-wave phase maps from sparse electrograms, even under substantial missing data.