Accelerating cardiovascular digital twin creation using precomputed virtual populations as Bayesian priors

Mechanistic lumped-parameter models of the cardiovascular system can provide insight into underlying physiological processes that cannot be measured directly. When tailored to an individual patient, these models have the potential to inform treatment for conditions such as heart failure. We aim to develop a robust and repeatable method of constructing a virtual population, defined as a collection of parameter sets that spans the full range of expected phenotypic variability of a population of interest. Such virtual populations can then accelerate patient-specific Bayesian inference by serving as a precomputed prior distribution to determine probabilistic cardiovascular digital twins.

This work presents a Bayesian approach to establishing a virtual population based on population-level measurement data. It utilizes the DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm to approximate a Bayesian posterior distribution consistent with the published clinical data of healthy populations (19 pressure and volume measurements across samples of at least 125 patients apiece). It subsequently applies weight-based allometric scaling of the resulting posterior parameter sets to more fully match target measurement variability. We applied this technique to a mechanistic hemodynamic model beginning with 23 parameters of interest. The final virtual population is then used to rapidly compute 17 proof-of-concept patient-specific digital twins, using patient data from an independent cohort not considered during virtual population development.

Our methodology effectively replicated at least 74.6% of the expected coverage of the target measurement distributions, with 17 of 19 measurements being represented at more than 80% of the expected coverage. Seventeen patient-specific digital twins were then computed from within this virtual population, with mean absolute percentage error across measurements for each patient ranging from 0.71% to 8.61% relative to the clinical measurement targets and a mean error across patients of less than 3%.

The proposed methodology represents a promising approach to developing a virtual population and demonstrates the feasibility of rapid posterior reweighting within such a precomputed virtual population, according to individual patient data. This computational proof-of-concept may help set the stage for future scalable creation of patient-specific digital twins to aid in clinical decision-making.