Enabling Real-Time Intervention Planning for Coronary Artery Disease using Machine Learning

Introduction: Machine learning (ML) assisted intervention planning can give physicians an intuitive approach to performing virtual interventions and receiving real-time hemodynamic feedback to guide clinical decision-making. Current FDA-approved tools for coronary intervention planning require days of processing and rely on conventional 2D displays for hemodynamic evaluation [1]. Long processing times prevent such tools from being applied in ambulatory cases with unstable coronary disease, such as ST-elevated myocardial infarction patients, where door-to-balloon time – the time from hospital admission to intervention – is recommended to be less than 90 minutes [2]. In addition, immersion offered by extended reality (XR) has been found to benefit intervention planning over traditional 2D displays. To our knowledge, no intervention planning tool exists with a turnaround time within one working day and is compatible with commonly available XR headsets.

To enable intervention planning with short turnaround times that leverage XR, we present HarVI (HARVEY [3] Virtual Intervention). HarVI is the back-end that predicts post-intervention hemodynamics. We focus on fractional flow reserve (FFR) [4] – a pressure-based metric that indicates coronary ischemia [5]. For user interaction, we use Harvis [6] – an established platform to modify geometries, deploy massively parallel simulations, and visualize hemodynamic results [6]. We introduce an innovative tool for intuitive clinical decision support by leveraging real-time flow predictions we establish with HarVI alongside the immersive anatomical editing capabilities from Harvis (Figure 1). This work lays the groundwork for developing immersive, intuitive, and real-time techniques for clinical decision support in coronary interventions.

Materials and

Methods: HarVI uses a clinically validated 1D computational fluid dynamics (CFD) framework [7] to train an ML model for real-time flow prediction (Figure 1A). Our analysis incorporated patients from two centers (n=75).

The training set for the ML model was constructed from a series of randomly sampled geometric modifications in terms of stenosis length and radius (Figure 1B). 1D CFD simulations were performed to obtain post-intervention FFR maps. We applied Latin Hypercube Sampling (LHS) to sample the parameter space efficiently. For each training instance (T_N), each 1D simulation was allocated one CPU, and all simulations were deployed simultaneously (Figure 2). Bivariate cubic splines were fitted for each point in the 1D geometry as a function of stenosis radii and lengths from the LHS matrix.

At this point, the one-shot training process (Figure 1A-B) would be completed. If the baseline FFR map shows locations with FFR < 0.80, an intervention would be required (Figure 3). To perform a virtual intervention, two dots are placed at the endpoints of a stenosis. Sliders allow the user to control the diameter of the stenosis, and the distance between the two dots adjusts the length of the stenosis. The corresponding stenosis length and radius specified by the user would be relayed from the Harvis GUI to the HarVI backend to query the resulting post-intervention FFR map. This planning stage can be performed any number of times without running any new 1D simulations (Figure 1C).

Results, Conclusions, and Discussions: We validated HarVI against 1D ground-truths for two test interventions. The correlation coefficient was 0.99 (p < 0.001) and the bias was −0.0006±0.0070 for Rx 1. The correlation coefficient was 0.97 (p < 0.001) and the bias was −0.0006±0.0200 for Rx 2. In short, FFR-HarVI almost perfectly recovers FFR-1D for continuous metrics with negligible bias in both test treatments. For context, FDA-approved software in the literature has reported a bias of around 0.01±0.05 [8, 4] when comparing FFR computed using CFD with invasively measured FFR.

The diagnostic performance of FFR-HarVI to discern stenoses with ischemia is summarized in Table 1. HarVI classified patients in full agreement with 1D CFD ground truths for both test cases except for those located right at the ischemic threshold. For Rx 2, only two cases were misclassified in the test set. Both instances had post-intervention FFR right at the 0.80 threshold, resulting in false negatives.

The clinical translatability of HarVI depends on the turnaround time required for the one-shot training process before real-time predictions can be made. For the HarVI end-to-end pipeline, reconstructions took 15-30 minutes, quality control to ensure adequate reconstruction required another 30 minutes, and the processing time for all patients was on average 13 minutes (Figure 4). The total turnaround time for HarVI was only 73 minutes on average, which fits in the door-to-balloon clinical standards of 90 minutes for patients with myocardial infarction [2]. After the one-shot training process, updating FFR-HarVI in response to the intervention took an average of 0.4 s for all centerline points using local machines.

In summary, we have taken initial steps toward integrating machine learning and extended reality for real-time planning of coronary interventions, focusing on focal lesions. HarVI offers a streamlined method for interventional cardiologists to evaluate treatment options in hours instead of days. The bivariate cubic-spline model used here demonstrates strong accuracy. Future exploration of other ML-based approaches could improve accuracy and reduce turnaround time. Although the developments presented here are in their early stages, they lay the critical foundation for improving the efficiency of clinical decision making in cardiology.

Acknowledgements (Optional): Computing support for this work came from the Duke Compute Cluster. The work of Cyrus Tanade was supported by the National Science Foundation Graduate Research Fellowship under Grant No. NSF GRFP DGE 1644868. The work of Amanda Randles was supported by the National Institute On Aging of the NIH under Award Number DP1AG082343 and the Duke/Duke-NUS Research Collaboration Pilot Project program. The content does not necessarily represent the official views of the NSF or NIH.

References:

[1] S. Sankaran, D. Lesage, R. Tombropoulos, N. Xiao, H. J. Kim, D. Spain, M. Schaap, and C. A. Taylor, “Physics driven real-time blood flow simulations,” Computer Methods in Applied Mechanics and Engineering, vol. 364, p. 112963, 2020.
[2] D. S. Menees, E. D. Peterson, Y. Wang, J. P. Curtis, J. C. Messenger, J. S. Rumsfeld, and H. S. Gurm, “Door-to-balloon time and mortality among patients undergoing primary pci,” New England Journal of Medicine, vol. 369, no. 10, pp. 901–909, 2013.
[3] 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.
[4] 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.
[5] P. A. Tonino, B. De Bruyne, N. H. Pijls, U. Siebert, F. Ikeno, M. vant Veer, V. Klauss, G. Manoharan, T. Engstrøm, K. G. Oldroyd, et al., “Fractional flow reserve versus angiography for guiding percutaneous coronary intervention,” New England Journal of Medicine, vol. 360, no. 3, pp. 213–224, 2009.
[6] H. Shi, J. Ames, and A. Randles, “Harvis: an interactive virtual reality tool for hemodynamic modification and simulation,” Journal of Computational Science, vol. 43, p. 101091, 2020.
[7] C. Tanade, S. J. Chen, J. A. Leopold, and A. Randles, “Analysis identifying minimal governing parameters for clinically accurate in silico fractional flow reserve,” Frontiers in Medical Technology, vol. 4, p. 1034801, 2022.
[8] 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.