Context: Computational models of cardiomyocytes are widely used to investigate disease mechanisms and drug effects. Increasing physiological accuracy, however, typically leads to greater model complexity, making parameter tuning computationally expensive and time-consuming. The selection of efficient optimization strategies is therefore becoming increasingly critical, yet it remains unclear which algorithms provide the best balance between convergence performance and computational cost.
Aim: This study compares the performance and computational cost of three optimization algorithms, the genetic algorithm (GA), particle swarm optimization (PSO), and sailfish optimization (SFO), for calibrating a coupled electro-chemo–mechanical cardiomyocyte model to represent hypertrophic cardiomyopathy (HCM) variants associated with myosin binding protein C mutations c.772G>A and Gln1061X.
Methods: Model parameters were selected based on known mutation-specific effects and Morris sensitivity analysis. Six contractile parameters were tuned for the c.772G>A variant and seven electrophysiological parameters for the Gln1061X variant. Each algorithm was run ten times using multi-objective formulations; six objectives for c.772G>A, and three for Gln1061X. Optimization was terminated after either 500 iterations or stagnation over 100 iterations. Performance was evaluated using runtime, convergence behavior, and spacing metrics. Model feasibility was further assessed against physiological biomarkers, both within and outside the objective function.
Results: All optimizations stopped due to stagnation, and 10-98% of Pareto solutions yielded plausible biomarkers. GA spacing was 67.9±28.5 and 98.2±4.6% lower than PSO or SFO spacing for c.772G>A and Gln1061X, respectively. PSO ran over 20% and 3.2% faster for Gln1061X and c.772G>A, respectively. The run times were dependent on the seed, with some runs converging faster using GA or SFO.
Conclusions: GA showed minor advantage in performance in solution diversity. However, the marginal differences in performance between algorithms suggest that careful objective design and parameter selection may have a greater impact on model calibration than the choice of optimization algorithm.