Assistant Professor University of California, Irvine Irvine, California, United States
Introduction:: Digital twin technology is an emerging concept in which a patient-specific diagnosis and/or therapy is based on using a virtual model (the digital twin) to predict and personalize treatment and prognosis for a patient (the real-life twin). Digital twins of the heart could improve treatment outcomes by providing optimized, patient-specific strategies for therapies such as cardiac resynchronization therapy, mitral valve repair versus replacement, and coronary artery bypass. While promising strides have been made in cardiovascular engineering to create digital twins of the heart, most current models: (1) predict short-term rather than long-term treatment outcomes, (2) are computationally expensive, and (3) do not consider data and model uncertainty. We will highlight our lab’s efforts to develop an efficient pipeline to build digital twins of the heart that can predict long-term cardiac remodeling in a real-time clinical setting using a combination of physics-based modeling and machine learning.
Materials and Methods:: We recently developed a rapid computational model of the heart and circulation that can predict long-term remodeling outcome following cardiac resynchronization therapy (CRT) (Figure 1) [1]. In brief, a lumped parameter model was used to model the circulation, and a simplified spherical representation of the cardiac walls was used to model atrioventricular mechanics. Cardiac growth was simulated using a strain-driven kinematic growth framework using fiber strain to drive isotropic growth of individual ventricular wall segments.
Classical techniques for calibrating computational models typically use iterative methods to find a single model parameter set that best matches the data. However, these methods require a high number of model evaluations, which is impractical for routine clinical use, and do not consider the uncertainty inherent to both model prediction and clinical data acquisition. We used a novel machine-learning approach based on Gaussian Process Emulation of model simulations using probabilistic surrogate models adapted from [2]. This method enables model parameter inference via a Bayesian History Matching technique. This method has two main advantages: (1) it uses surrogate models that can be evaluated at extremely low computational cost and (2) result in probabilistic model outcomes that consider both model and data acquisition uncertainty. Bayesian history matching has been used to calibrate complex models ranging from galaxy formation to SARS-CoV2 impact, and only recently on cardiac electrophysiology and mechanics models [2, 3]. We here make this method more computationally efficient by coupling it to our fast model of the heart and circulation.
Results, Conclusions, and Discussions:: We first tested our model personalization pipeline on a synthetic data set, i.e. simulations results using known model parameters, where we assumed an uncertainty of 10% of the “observed” values according to [4]. The history matching method successfully identified a range of parameters for which the model outcomes were within the synthetic data uncertainty (Figure 2a-c). The original, known parameter values were encompassed by the fits. The additional advantage of using computationally efficient emulators is fast performance of sensitivity analysis, which we used to select only the most important parameters to be included in the fitting process (Figure 2d).
So far we have tested our digital twin framework for two patients suffering from non-ischemic left bundle branch block treated with CRT: one responder and one non-responder. Routine clinical data (including 12-lead ECG, MRI, cuff blood pressure) was used to personalize the model using the history matching approach and 6 months of post-CRT remodeling was predicted for each potential lead placement location. Interestingly, our model predicted that both patients could have had a more beneficial degree of reverse remodeling, here shown as mean 6-month change in left ventricular end-diastolic volume (EDV), with a different lead location selected (Figure 1). Most importantly, CRT for the non-responding patient could have significantly reduced EDV when pacing from a different wall segment, thus resulting in a beneficial CRT response (see figure).
We will continue training and testing our physics-based modeling and machine learning digital twin pipeline to a larger population of patients treated with CRT to further increase our confidence that this model can predict the reverse remodeling that occurs during CRT in individual patients. We will include ischemic and nonischemic heart failure patients to determine how ischemia affects prediction accuracy. This work is an important step in creating a digital twin of the heart that can identify the best CRT treatment strategy for each individual patient before clinicians reach the operating room.
Acknowledgements (Optional): : The work that will be presented entails the results of various collaborations with partners at the University of California Irvine, University of Virginia, University of Alabama at Birmingham, and Siemens Healthiness. This work was supported by funding from the National Institutes of Health (R01HL159945).
References (Optional): : [1] Oomen, P. J. A., Phung, T.-K. N., Weinberg, S. H., Bilchick, K. C. & Holmes, J. W. A rapid electromechanical model to predict reverse remodeling following cardiac resynchronization therapy. Biomech Model Mechan 21, 231–247 (2022). [2] Coveney, S. & Clayton, R. H. Fitting two human atrial cell models to experimental data using Bayesian history matching. Prog Biophysics Mol Biology 139, 43–58 (2018).
[3] Longobardi, S. et al. Predicting left ventricular contractile function via Gaussian process emulation in aortic-banded rats. Philosophical Transactions Royal Soc 378, 20190334 (2020).
[4] Kou, S. et al. Echocardiographic reference ranges for normal cardiac chamber size: results from the NORRE study. European Hear J - Cardiovasc Imaging 15, 680–690 (2014).