Introduction: Personalisation of cardiac electrophysiology models remains an important challenge for development of digital twins that can be used to guide interventions such as radiofrequency ablation to treat atrial fibrillation. The focus of this work is on global sensitivity analysis, which indicates how model parameters affect behaviour and is an important step towards calibration of models using data that are accessible in the clinical setting.
Methods: We used the monodomain equation with excitability described by the modified Mitchell-Schaeffer (mMS) model with isotropic diffusion and homogenous parameter fields. Tissue geometry of the left atrium (LA) was estimated in 12 different patients by segmentation of CMR images. An input dataset of 202 samples of 5 mMS model parameters was generated using Latin hypercube sampling. These samples were used to run 202 simulations in each geometry, with 3xS1 beats at 800ms cycle length followed by an S2 beat with a coupling interval of 500ms. Local activation time (LAT) and action potential duration (APD) were determined across the mesh. These outputs were then used to build Gaussian process emulators of LAT and APD as a function of the 5 mMS model inputs, for each geometry. The emulators were used to determine main effect and total effect sensitivity indices using variance based global sensitivity analysis.
Results: The LAT and APD were most sensitive to tau_in, tau_out and diffusion, with these three parameters explaining more than 95% of the variance in the outputs when varying one input at a time. The overall sensitivity indices were similar across geometries.
Conclusion: LAT and APD depend principally on three mMS model parameters, and that this dependence is not strongly affected by geometry. This finding indicates that these parameters can in principle be recovered from clinical data.