Using personalized computer models to custom-tailor ablation procedures for atrial fibrillation patients: are we there yet?

KEYWORDS:

• Atrial fibrillation
• catheter ablation
• computational modeling
• fibrotic remodeling
• late gadolinium-enhanced magnetic resonance imaging

Atrial fibrillation (AF) is the most common sustained heart rhythm disorder, affecting 1–2% of the global population and contributing significantly to mortality and morbidity rates [1]. Some forms of this arrhythmia can be successfully treated via catheter ablation, but the success rate of such procedures for individuals with persistent AF (PsAF) is dismal (~50%) [2]. This is attributed in part to the presence of extensive atrial fibrotic remodeling in PsAF patients, which confounds strategies for identifying ablation targets [3]. This problem is exacerbated by significant inter-patient variability in the spatial distribution of fibrotic tissue [4]. As such, there is a pressing need to devise new ablation strategies that incorporate understanding of how each PsAF patient’s pattern of fibrotic remodeling contributes to arrhythmia perpetuation.

A potential means of addressing this challenge has emerged from the domain of computational modeling of cardiovascular physiology. This research area has enjoyed robust growth since the late twentieth century to the point where computational approaches are now established as an important means of investigating unexplored clinical questions in many cardiac fields, including hemodynamics [5,6], electromechanics [7], optogenetics [8], and electrophysiology [9,10]. With the advent of late gadolinium enhancement magnetic resonance imaging (LGE-MRI), which can be used to ascertain the distribution of fibrotic tissue in each individual’s atria [11], it is now possible to reconstruct computational models of the atria of PsAF patients that incorporate each individual’s geometrically detailed anatomy and spatial pattern of structural remodeling. Recent publications from our lab have described how such models can be used to analyze mechanistic underpinnings of reentrant drivers (RDs) that are thought to help perpetuate PsAF [12] and explore new strategies for personalizing atrial ablation procedures [13].

Several key factors differentiate the latter studies from earlier work in the realm of computational modeling of the atria, which we recently reviewed in detail [10]. First, unlike many previous modeling projects that considered only the left atrium (LA), our most recent reconstructions are bi-atrial, which is important because RDs are observed in both chambers. Second, both studies involved a large number of patients (n = 20 [12] and n = 10 [13], respectively), each of whom had a unique pattern of structural remodeling. To the best of our knowledge, the largest prior computational study of the fibrotic atria to these used four LGE-MRI-based models of the LA only [14–16]. Third, we used a novel methodology in which computational models were carefully calibrated to realistically represent the electrophysiological properties of fibrotic and nonfibrotic atrial tissue in PsAF [12]. Pathophysiological effects of PsAF on the action potential in fibrotic regions were incorporated by modifying ionic currents based on in vitro observations of atrial myocytes subjected to elevated high levels of transforming growth factor β1. This approach allowed us to build experimentally constrained models of the fibrotic atria [10].

Finally, in both studies, significant efforts were made to show correspondence between organ-scale model behavior and relevant clinical observations in the same patients whose LGE-MRI scans were used for model reconstruction. In the study concerned with PsAF [12], we mapped trajectories of RD organizing centers recorded via electrocardiographic imaging into the corresponding patient models and showed the majority (56.7 ± 9.1%) were located in a small subset of atrial tissue (13.79 ± 4.93% of overall atrial volume) in which our analysis identified a fibrosis spatial pattern that strongly favored dynamic localization of RDs in simulations. Our second study was concerned with atypical LA flutter [13], a common form of post-ablation recurrence in AF patients. Here, we adapted an algorithm from graph theory to predict optimal ablation lesions for rendering arrhythmia initiation impossible in personalized atrial models. Ablation targets derived via this approach abolished LA flutter inducibility in silico and were located in the same atrial regions as clinical ablation lesions that successfully terminated arrhythmia in the corresponding patients.

It should be noted that other studies carried out contemporaneously with those discussed earlier proposed alternative approaches for representing fibrotic tissue in personalized atrial models. For example, Bayer et al. [17] used a probabilistic approach based on LGE-MRI data from several PsAF patients to calibrate electrophysiological parameters in a single LA bi-layer model. This platform was used to explore novel several ablation strategies, including an effective approach where lesions followed streamlines that tracked the activation sequence during sinus-like activation. This method and/or other alternative strategies for modeling the fibrotic substrate in PsAF, as recently reviewed by Roney et al. [18], might be used instead of or in combination with our approach (described earlier) if this proves to be beneficial for planning PsAF procedures.

To this point, we have established that state-of-the-art computational modeling tools can simulate arrhythmia episodes that are consistent with clinical observations from PsAF patients and that analysis of simulation results can yield information that might be relevant for the planning of ablation procedures. In the remainder of this editorial, we outline how these tools could be assembled to carry out custom-tailored PsAF ablation procedures, with particular attention paid to potential hurdles and how they might be overcome.

The first step of the approach envisioned by our lab is the reconstruction of a 3D geometric model of the atria, which begins with acquisition of LGE-MRI scans from a patient with PsAF. For our two previous studies [12,13], imaging was conducted at two different centers (University of Bordeaux and Johns Hopkins Hospital, respectively), suggesting that the same approach could be used by other groups. We have found that LGE-MRI at standard clinical resolution (1.25 mm × 1.25 mm in-plane, 2.5 mm out-of-plane) produces images adequate for reconstructing models with strong predictive value [12,13], as long as patients are in sinus rhythm during the procedures. For patients who present in AF, pre-MRI cardioversion to achieve temporary restoration of sinus rhythm is necessary. Next comes image segmentation, during which the geometric contours of the left and right atria are identified using a semi-automatic approach and fibrotic tissue is differentiated from nonfibrotic regions using a normalized image intensity ratio method described previously [19]. A finite element mesh appropriate for accurate simulation of electrophysiological properties is then generated from segmented images using an automated approach [12,13]. Finally, fiber orientations are estimated via systematic deformation of an atlas with known fibrous structure [9], since it is not yet feasible to acquire such data via in vivo scans. In our experience, the approximate turnaround time between image acquisition and completion of model construction ranges from 8 to 16 h, with much of this time spent waiting for computationally intensive meshing and fiber mapping processes to complete.

Notably, the model reconstruction pipeline described here is highly modular, and the different steps can be individually upgraded to improve or speed up the process as new methods become available. For example, acquisition of higher-resolution atrial images would provide finer-grain anatomical detail that could potentially unveil additional insights on the arrhythmia substrate [20]. Likewise, estimation of atrial fiber orientations could be accelerated by developing a rules-based approach (as used in ventricular model development [21]) or greatly improved by drawing on multiple sources of high-resolution atrial fiber data [22] instead of a single atlas.

The second step in our proposed approach is to perform computational simulations in which the model is subjected to virtual programmed electrical stimulation (PES) from at least 30 sites distributed evenly throughout the left and right atria. As shown previously [12], this approach is sufficient for identifying regions where the spatial pattern of fibrotic tissue favors dynamic localization of RDs; in our view, these sites should be specifically targeted for ablation in order to render the patient-specific fibrotic substrate noninducible for AF. The targets produced by the initial round of simulations should then be virtually ablated by treating the local tissue volume as inexcitable and nonconductive, as in previous studies [23]. Then, a second round of virtual PES simulations should be conducted to pinpoint any emergent post-ablation RD sites, which have been reported elsewhere [24]. If such regions exist, they should also be targeted for ablation. The latter step highlights an important potential advantage of the proposed approach since it is capable of identifying all potential RD targets, not just those manifested prior to the ablation procedure. In our view, this could significantly reduce the risk of post-ablation AF recurrence.

Notably, access to high-performance computing resources is a pre-requisite for running the simulations described earlier. For example, we recently reported that simulation of 1000 ms of activity in a typical organ-scale cardiac model required 1 h 40 m of compute time on the largest and fastest grid computing resource available at our university (the Maryland Advanced Research Computing Centre) [25]. Based on this experience and assuming that simulations for different pacing sites can be run in parallel, we estimate that each round of virtual EPS experiments for custom-tailored PsAF ablation procedures will take 8–12 h.

The third and final step in our proposed approach is transfer of recommended RD ablation targets into a reduced resolution geometric mesh that can be loaded into electro-anatomical mapping systems commonly used to guide AF ablation procedures (e.g. CARTO-Merge, Biosense-Webster). Since the specific purpose of ablating these target locations is to eliminate the propensity of the fibrotic substrate to initiate and perpetuate RDs, the corresponding lesions can be created either before or after any other ablation lesions (e.g. to isolate the pulmonary veins) as part of the standard of care for AF.

Based on the details provided earlier and factoring a generous margin of error for simulation analysis, data transfer, troubleshooting, and so on we anticipate that it should be possible to complete all steps detailed earlier for a single patient within 2.5–5 days. In our experience, patient LGE-MRI scans often occur 7–10 days prior to ablation procedures. As such, the proposed approach is compatible with clinically relevant timelines for PsAF patients.

In conclusion, we are confident that custom-tailored planning of PsAF ablation procedures based on noninvasive analysis of personalized computational models is feasible, and we expect proof-of-concept testing to begin in the near future. If successful, we anticipate that this approach will improve ablation success rates, make procedures simpler and less time-consuming, and ensure long-term freedom from AF.

Declaration of interest

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Additional information

Funding

This work was supported by American Heart Association (grant number 16-SDG-30440006, to P.M. Boyle), by an NSF Graduate Research Fellowship and an ARCS Foundation Award (to S. Zahid), and by the Johns Hopkins Medicine Discovery Fund and National Institutes of Health (grant number DP1-HL123271, to N.A. Trayanova).