Introduction and Aims: To model variability of cardiac action potentials (APs), a population of models consisting of different sets of a model's parameter values can be created and calibrated to match observed variability in properties such as AP duration (APD). However, identifying appropriate parameter sets for the virtual cohort can be difficult and time-consuming.
Methods: We adapt a particle swam optimization (PSO) optimization technique to generate a population of models. Our population PSO (PPSO) algorithm discourages convergence to a local minimum, and instead guides the search to explore low-error areas of parameter space, yielding many parameter sets that reproduce the variability of biomarkers seen in real tissue data.
Results: Using canine ventricular action potential (AP) microelectrode recordings, we extracted a set of biomarkers consisting of APDs measured at 90\% and 50\% repolarization as well as peak voltage. Variability was represented by allowing narrow (±10\%) and wide (±30\%) variation of the base biomarker values. We created a 5000-member population of models fitting the 13 parameters of the Fenton-Karma (FK) model to the biomarker ranges using PPSO and compared it with a random approach; in both cases, the first 5000 candidate parameter sets that produced APs within the biomarker ranges were accepted to form the population (see panel A). PPSO accepted 54\% and 21\% of the candidate parameter sets for the wide and narrow tolerances, respectively, compared to 18\% and 2.3\% with the random approach, while retaining similar variability in parameter values (see panel B).
Conclusion: Compared to a random approach, our novel PPSO method can generate populations of models matching biomarkers more efficiently, accepting 3-10 times more candidate parameter sets when fitting biomarkers from canine AP recordings to the FK model.