(M-101) Cardiovascular Model Parameterization of Digital Twins for Acute Care Drug Delivery: An Iterative Bayesian Approach

Jon Peterson, PhD – Distinguished Scientist, Medical and Health Informatics Laboratories, NTT Research, Inc.; Giulia Cornali, MS – Research Scientist, Medical and Health Informatics Laboratories, NTT Research, Inc.; Jennifer Rohrs, PhD – Co-founder, Director of Modeling, Embody Biosciences; Patrick Gelbach, PhD – Research Scientist, Medical and Health Informatics Laboratories, NTT Research, Inc.; Joe Alexander, MD, PhD – Senior Vice President and Director, Medical and Health Informatics Laboratories, NTT Research, Inc.

Objectives: A mechanistic model of the cardiovascular system can represent blood flow throughout the circulatory system under healthy and pathological conditions [1]. It is well suited for use as the foundation for an autonomous closed-loop intervention system (ACIS), which would deliver multiple drugs simultaneously based in part on bio digital twin predictions of patient-specific drug responses. The underlying hemodynamic model includes numerous parameters that have biological relevance but cannot be directly measured in a clinical setting, thus requiring a method of parameter estimation that can accommodate a large parameter space. This work presents an iterative process consisting of single-parameter sensitivity analysis and multi-parameter Bayesian inference to determine a parameter set consistent with population-level clinical data.

Methods: The process begins with single-parameter sensitivity analysis to identify parameter ranges such that each input parameter creates a target amount of variability in at least one output measurement and each output measurement has its target variability created by at least one input parameter. These ranges are then used as bounds for Latin Hypercube Sampling to create a digital population of parameter sets with all parameters varying simultaneously (the Bayesian prior). Bayesian inference using population-level pressure and volume data (the Bayesian evidence, from [2-4]) then identifies parameter sets most likely to explain the observed measurements, outputting posterior probability distributions for each parameter. The median of each distribution can then be used to start the next round through the iterative process.

Results: After 6 rounds of the process, normalized root-mean-square error (normalized to each measurement’s standard deviation from literature) decreased from approximately 0.66 with initial parameter values to 0.12 with the parameter medians from the final round, with all measurements falling within 0.27 standard deviations of their target mean.

Conclusions: The iterative sensitivity analysis and Bayesian inference process is a promising tool to estimate representative parameter values for a healthy patient population. These techniques can be adapted to find representative parameter sets for other populations, including different body sizes or disease states, supporting the development of a digital population from which to select patient-specific digital twins to aid in acute care drug delivery when treating cardiovascular disease.

Citations: [1] Shimizu, S. et al. Lumped parameter model for hemodynamic simulation of congenital heart diseases. J Physiological Sci 68, 103–111 (2018)
[2] Kawel-Boehm, N. et al. Reference ranges (“normal values”) for cardiovascular magnetic resonance (CMR) in adults and children: 2020 update. J Cardiov Magn Reson 22, 87 (2020)
[3] Wright, J.D. et al. Mean Systolic and Diastolic Blood Pressure in Adults Aged 18 and Over in the United States, 2001–2008. National Health Statistics Reports (2011)
[4] Kovacs, G. et al. Pulmonary arterial pressure during rest and exercise in healthy subjects: a systematic review. Eur. Respir. J. 34, 888–894 (2009)