Enabling Wearable-Driven Longitudinal Hemodynamic Maps of 3D Blood Flow using Cloud Computing

Introduction:: A digital twin of a patient’s cardiovascular arteries can accurately capture hemodynamic responses to physiological changes [1, 2]. However, 3D coronary simulations are conventionally limited to single heartbeat metrics due to their high computational cost [1, 2, 3]. Indeed, even simulating one day of activity using current methods would take hundreds of years on the world’s largest supercomputers. As a result, monitoring vascular responses to stimuli and treatment over time remains a grand challenge. To overcome this challenge, we propose the Longitudinal Hemodynamic Mapping Framework (LHMF), a multiscale approach that combines 3D physics-based digital twins, wearable devices, and cloud computing to monitor hemodynamic changes over substantial periods of time.

To expand our understanding of how hemodynamics evolve longitudinally and to identify disease-prone areas, we created the first 3D patient-specific Longitudinal Hemodynamic Maps (LHMs). LHMs capture the full range of behavior over one week, unlike conventional hemodynamic maps that show only one snapshot in time of a patient’s condition. LHMs quantify how blood flow develops over time and provide insights into whether specific anatomical locations predispose to disease. We enable LHMs by developing methods to 1) drive simulations from continuous wearable data, 2) efficiently utilize high-throughput computing resources, and 3) extend the temporal domain over which blood flow is measured. We establish and validate the LHMF that enables longitudinal hemodynamic assessment, thereby moving one step closer to improved risk stratification in treatment planning and recovery, such as predicting hemodynamic responses to bypass surgery or stent placement months post-procedurally.


Materials and Methods:: Hemodynamic simulations were performed using our in-house fluid solver HARVEY [4, 5]. Transient velocity waveforms were coupled to continuous wearable data (Fig. 1A) at the inlet boundary condition (Fig. 1B). Heart rate data informed cardiac period and amplitude, and electrocardiography tracings informed systolic and diastolic durations. We focused our hemodynamic assessment on wall shear stress (WSS) - the frictional force of blood acting on the inner walls of blood vessels.

We established the LHMF, enabling the decomposition of a sequence of heartbeats into set hemodynamic units. A hemodynamic unit consisted of capturing preceding cardiac cycles for temporal convergence and the heartbeat of interest. Rather than simulating heartbeats contiguously – the established approach, our framework analyzed hemodynamic units in parallel on cloud resources (Fig. 1C). Stitching results of all hemodynamic units was equivalent to running a contiguous simulation using established methods (Fig. 1D). To validate the framework, we conducted a 750 heartbeat simulation using the established approach and compared WSS with the results obtained using LHMF. We refined LHMF by clustering hemodynamically similar heartbeats, establishing LHMF-C. We identified clusters of similar input conditions such that the resulting 3D metrics were not statistically different. Hemodynamically similar outputs were measured using Bland-Altman bias and imprecision, which has been used extensively to validate computational hemodynamic models with in vivo measurements. We tolerated a bias of 0.01 and imprecision of 0.05 to match FDA-approved software [2, 6]. We applied LHMF-C to evaluate WSS continuously for one week of activity on 38,400 CPUs of cloud resources.


Results, Conclusions, and Discussions:: Comparing LHMF to established methods, WSS had a percent error of 3.9x10^-3 % – attributable to precision error. This validation demonstrated that LHMF reconstructs explicit simulations accurately. We improved the tractability of LHMF via clustering – establishing LHMF-C.

To test LHMF-C, we analyzed 710 thousand heartbeats (1 week of activity). We found that 680 hemodynamic units were required to reconstruct the full range of activity, demonstrating that clustering improved the tractability of performing longitudinal simulations. Simulating all the hemodynamic units required 25.8 hours when neglecting overhead. Since low WSS (< 1 Pa) is linked to the development of atherosclerotic plaque [7], we focused on generating an LHM for the first time that displays the percentage of time with low WSS, providing a novel way to identify local regions with potential risks for disease development. Low WSS frequently occurred in the stenotic region during the week, depicted in the red rectangular inset in Fig. 1D. Prolonged exposure to low WSS can cause endothelial cell dysfunction and has been linked to the development of plaque and hypertension [8]. Beyond identifying sites of low WSS, measuring how long certain locations persist with low WSS is crucial to better stratify potential risks to cardiovascular health.

To capture hemodynamic nuances that can only be observed over time, we compared conventional WSS maps averaged over single heartbeats (indicated by black insets in Fig. 1D) with the LHM. The rest state exhibited a larger area of low WSS, while the exercise state had minimal regions of low WSS compared with the LHM. Our results suggest that LHMs can capture more comprehensive hemodynamic information than single heartbeat maps.

In conclusion, we present the first critical steps in enabling analysis of longitudinal hemodynamics by introducing and validating the LHMF. Future work includes tuning the model inputs from consumer-grade wearable devices and using machine learning to minimize the number of required hemodynamic units. Ultimately, we expect the LHMF to have wide ranging impact beyond the study of cardiovascular disease by shifting the use of vascular digital twins from several heartbeat studies to models capturing changes over weeks to months.


Acknowledgements (Optional): : Computing support for this work came from Amazon Web Services (AWS) Cloud Credits for Research, Duke Compute Cluster, and Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) which is supported by NSF grant number ACI-1548562. An award of compute time was provided by the INCITE program. This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DEAC05-00OR22725. This work was supported by the AWS team through architecture reviews to optimize for the cloud as well as providing access to high-performance compute cores. Research reported in this work was supported by the National Institute On Aging of the NIH under Award Number DP1AG082343, NSF GRFP under Grant No. DGE 164486, and the Coulter Foundation. The content does not necessarily represent the official views of the NIH or NSF.


References (Optional): :

1. 
   M. Vardhan, J. Gounley, S. J. Chen, E. C. Chi, A. M. Kahn, J. A. Leopold, and A. Randles, “Non-invasive characterization of complex coronary lesions,” Scientific reports, vol. 11, no. 1, pp. 1–15, 2021.
   
   


2. 
   B. L. Nørgaard, J. Leipsic, S. Gaur, S. Seneviratne, B. S. Ko, H. Ito, J. M. Jensen, L. Mauri, B. De Bruyne, H. Bezerra, et al., “Diagnostic performance of noninvasive fractional flow reserve derived from coronary computed tomography angiography in suspected coronary artery disease: the nxt trial (analysis of coronary blood flow using ct angiography: Next steps),” Journal of the American College of Cardiology, vol. 63, no. 12, pp. 1145–1155, 2014.
   
   


3. 
   M. Vardhan, J. Gounley, S. J. Chen, A. M. Kahn, J. A. Leopold, and A. Randles, “The importance of side branches in modeling 3d hemodynamics from angiograms for patients with coronary artery disease,” Scientific reports, vol. 9, no. 1, pp. 1–10, 2019.
   
   


4. 
   A. P. Randles, V. Kale, J. Hammond, W. Gropp, and E. Kaxiras, “Performance analysis of the lattice boltzmann model beyond navier-stokes,” in 2013 IEEE 27th International Symposium on Parallel and Distributed Processing, pp. 1063–1074, IEEE, 2013.
   
   


5. 
   A. Randles, E. W. Draeger, T. Oppelstrup, L. Krauss, and J. A. Gunnels, “Massively parallel models of the human circulatory system,” in Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pp. 1–11, 2015.
   
   


6. 
   M. Pellicano, I. Lavi, B. De Bruyne, H. Vaknin-Assa, A. Assali, O. Valtzer, Y. Lotringer, G. Weisz, Y. Almagor, P. Xaplanteris, et al., “Validation study of image-based fractional flow reserve during coronary angiography,” Circulation: Cardiovascular Interventions, vol. 10, no. 9, p. e005259, 2017.
   
   


7. 
   H. Samady, P. Eshtehardi, M. C. McDaniel, J. Suo, S. S. Dhawan, C. Maynard, L. H. Timmins, A. A. Quyyumi, and D. P. Giddens, “Coronary artery wall shear stress is associated with progression and transformation of atherosclerotic plaque and arterial remodeling in patients with coronary artery disease,” Circulation, vol. 124, no. 7, pp. 779–788, 2011.
   
   


8. 
   E. Cecchi, C. Giglioli, S. Valente, C. Lazzeri, G. F. Gensini, R. Abbate, and L. Mannini, “Role of hemodynamic shear stress in cardiovascular disease,” Atherosclerosis, vol. 214, no. 2, pp. 249–256, 2011.