Patient-specific cardiac models can be used to improve the diagnosis of cardiovascular diseases. However, practical application of these models is impeded by the computational costs and numerical uncertainties of fitting them to clinical measurements from individual patients. Reliable and efficient model tuning within medically-appropriate error bounds is a requirement for practical deployment in the time-constrained environment of the clinic. In this work, we present a framework to efficiently tune parameters of patient-specific mechanistic models using routinely acquired non-invasive patient data with a hybrid particle swarm and pattern search optimization algorithm.

The proposed framework is used to tune full-cycle lumped parameter circulatory models using clinical data obtained from patients as well as canine subjects; showing that the framework can be easily adapted to optimize cross-species models. It is also used to simultaneously obtain the unloaded geometry and passive myocardial material parameters of four left-ventricular cardiac finite element models constructed from canine subject MRI data. This demonstrates that the proposed approach can support the use of complex models to obtain data that cannot be directly measured. The patients gave informed consent and the canine subject studies were approved by the local Institutional Review Boards. The optimized results in all case studies were within acceptable error tolerances.

Additionally, the framework is extended to include uncertainty quantification -- supporting model tuning with often-unreliable data sources that are ill-suited to a deterministic approach. The proposed approach for probabilistic model tuning discovers distributions of model inputs which generate target output distributions. Probabilistic sampling is performed using a model surrogate for computational efficiency and a general distribution parameterization is used to describe each input. The approach is tested on four test cases using CircAdapt, a cardiac circulatory model. Three test cases are synthetic, aiming to match the output distributions generated using known reference input data distributions, while the fourth example uses real-world patient data for the output distributions to obtain the input distribution. The results demonstrate accurate reproduction of the target output distributions, with accurate recreation of reference inputs for the three synthetic examples.

Overall, this work automates the use of biomechanics and circulatory cardiac models in both clinical and research environments by ameliorating the tedious process of manually fitting model parameters and supports the use of more complex models in practice through the quantification of error.