Nanoscale cardiac junctions no longer stall tissue simulations: a hybrid model runs a million junctions and reproduces conduction block
Synopsis
Cardiac tissue models have long collapsed each cell-to-cell junction into a fixed conductance, because finite-element models that resolve the nanoscale intercalated disk are confined to cables of tens of cells. A hybrid model that replaces each junction with a memory-based neural network flow map reproduces action potentials (1–2 mV RMSE) and conduction-velocity restitution (mean error 0.24%) in cables, and scales without retraining to 3-D tissue with 1.2 million junctions, 104–105 times faster than the finite-element model. In tissue with widened clefts and reduced gap-junctional coupling, it produces regional conduction block, whereas either change alone does not.
Interpretation
Cell-to-cell coupling can now enter whole-tissue cardiac simulations as a learned element instead of a fixed conductance that discards intercalated-disk nanostructure. The earlier finite-element model resolved hundreds of coupled cleft potentials, concentrations and membrane patches per junction, and that cost confined it to cables of a few tens of cells, out of reach of the 105–106 junctions a 2-D tissue needs to sustain reentry. In a 50-cell cable the surrogate matched the finite-element reference with 1–2 mV action-potential RMSE, and conduction velocity of the last beat was within 0.3% (strong coupling 36.98 versus 37.07 cm/s; reduced coupling 18.27 versus 18.31 cm/s).
Each junction is replaced by a flow map with memory that predicts the two cleft currents flowing into the adjacent cells from recent history alone. Flow map learning does not learn a differential-equation right-hand side; it learns the fixed-step advance of the state directly, so it avoids the stiffness of cleft dynamics and needs no equations for the observed quantities. The network takes 1 ms of cleft-current history plus the axial potentials and sodium-channel availability of both cells, 310 inputs and 2 outputs; across twelve combinations of six disk geometries and two couplings, the mean absolute conduction-velocity error over 178 measurement points was 0.24%, with a maximum of 1.18%.
A trained junction model drops into larger tissue unchanged, without retraining for tissue size. Models that resolve the disk had stayed at cable scale, so junction-to-junction heterogeneity in 2-D and 3-D tissue could not be tested. In a 100×100 sheet (19,800 junctions) and a 100×100×40 block (1,182,000 junctions) the surrogate ran about 104–105 times faster than the finite-element model and within an order of magnitude of a monodomain model that omits the disk; one beat of the 3-D block was estimated at about two and a half years with the finite-element model and took about 3 minutes with the surrogate.
Perspective
The demonstrated scope is point-by-point comparison with the finite-element model in homogeneous and heterogeneous 50-cell cables, validation on a 20×20 sheet, and demonstration of conduction block from regional cleft widening combined with reduced coupling in a 100×100 sheet and a 100×100×40 block; the 3-D tissue predictions remain to be tested against experiments.
Each surrogate inherits the approximations of the finite-element cable used to train it, in which all disk-localized membranes sit on the two cell ends, a geometry that does not strictly hold in higher dimensions because cleft currents depend nonlinearly on disk geometry and channel density. Real myocardial disks are not arranged on a Cartesian lattice, branched myocytes bear several disks of different size and structure, and gap junctions lateralize in disease.
