Physics-guided deep learning inference of intramural slow-conduction depth from simulated epicardial optical mapping

Abstract Epicardial optical mapping integrates fluorescence over depth, making intramural slow conduction difficult to localize from surface signals alone. We present a simulation-based proof-of-concept study coupling monodomain electrophysiology and Monte Carlo photon transport to generate epicardial optical action potentials (OAPs) from slow-conduction layers in a 5 mm ventricular wall. In an initial 71-case depth sweep with fixed layer properties, 5-fold cross-validation of a one-dimensional residual network gave $$r=0.963$$ , $$R^2=0.916$$ , and RMSE $$=0.298$$ mm, with reduced performance beyond approximately 3 mm. We then varied start depth (0.5–3.5 mm), thickness (0.15–0.50 mm), and conductivity multiplier (0.15–0.4) in a 279-case factorial study. Training the same architecture on this expanded dataset and using grouped-depth cross-validation, in which every substrate variant at each test depth was withheld, retained predictive depth information across all tested combinations ( $$r=0.852$$ , $$R^2=0.717$$ , RMSE $$=0.476$$ mm), although accuracy varied with layer thickness and conductivity. Previously acquired rabbit optical and depth-stepped microelectrode recordings showed behaviour consistent with the simulated signals but were not used to train or quantitatively validate the estimator. These results demonstrate that simulated OAP morphology contains recoverable information about slow-layer depth across the tested variations in layer thickness and conductivity.