Comparing Inducibility of Reentrant Arrhythmia in Patient-Specific Computational Models to Clinical Atrial Fibrillation Phenotypes

Abstract

Objective:

Test the hypothesis that conduction velocity (ϴ) modulation in persistent AF models can improve simulation agreement with clinical arrhythmias.

Background:

Computational models of fibrosis-mediated, reentrant left atrial (LA) arrhythmia can identify possible substrate for persistent atrial fibrillation (AF) ablation. Contemporary models use a one-size-fits-all approach to represent electrophysiological properties, limiting agreement between simulations and patient outcomes.

Methods:

Persistent AF patients (n=37) underwent ablation and were followed ≥2 years to determine post-ablation outcomes: AF, atrial flutter (AFL), or no recurrence (NR). Patient-specific LA models (n=74) were constructed using pre- and ≥90 days post-ablation MRI. Simulated pacing gauged in silico arrhythmia inducibility due to AF-like rotors or AFL-like macro-reentrant tachycardias (MTs). A physiologically plausible range of ϴ values (±10 or 20% vs. baseline) was tested and model/clinical agreement was assessed.

Results:

Fifteen (41%) patients recurred with AF and 6 (16%) with AFL. Arrhythmia was induced in 1078/5550 simulations. Using baseline ϴ, model/clinical agreement was 46% (34/74 models), improving to 65% (48/74) when any possible ϴ value was used (p=0.014, McNemar’s test). ϴ modulation improved model/clinical agreement in both pre- and post-ablation models. Pre-ablation model/clinical agreement was significantly greater for patients with extensive LA fibrosis (>17.2%) and elevated BMI (>32.0 kg/m²).

Conclusions:

Simulations in persistent AF models show a 41% relative improvement in model/clinical agreement when ϴ is modulated. Patient-specific calibration of ϴ values could improve model/clinical agreement and model usefulness, especially in patients with higher BMI or LA fibrosis burden. This could ultimately facilitate better personalized modeling, with immediate clinical implications.

Condensed Abstract

Computational modeling of reentrant arrhythmia in patients with persistent atrial fibrillation is feasible using image-derived data on anatomy and fibrotic substrate. Models incorporate changes to electrophysiological properties due to fibrosis but use the same default parameterization regardless of patient-specific variability, which may limit model agreement with clinical observations. Executing simulations with a range of conduction velocity values produced a set of possible predictions with improved model/clinical agreement compared to simulations that were run with default conduction velocity only.

Personalization of conduction velocity could thus significantly improve model/clinical agreement, especially in patients with increased left atrial fibrosis or body mass index.

1. Introduction

Ablation is a keystone of treatment for atrial fibrillation (AF), but results fall short especially in patients with persistent AF.¹ Structural and electrical remodeling, which lead to LA fibrosis, are strongly related to recurrence risk.^{2, 3} Computational simulations of reentrant arrhythmia in three-dimensional (3D) LA models have emerged as a powerful tool to characterize the spatial distribution of regions with the propensity to sustain arrhythmia and identify targets for ablation to reduce the risk of recurrent AF.^4–6

Computational LA models incorporate personalized (patient-specific) anatomy and fibrosis derived from late gadolinium enhancement magnetic resonance imaging (LGE-MRI). Yet, models incorporate standardized (non-personalized) changes to cell- and tissue-scale electrophysiology.⁷ For example, the same reduction in conduction velocity (ϴ) is applied to all fibrotic areas based on the observed effects of transforming growth factor beta-1 (TGF-β1) on rat cardiomyocytes,⁸ despite known differences in fibrotic biomarkers between patients.^9–11 Simulation of reentrant arrhythmia involves rapid pacing to assess whether a model sustains an electrical wavefront after the cessation of pacing.^{7, 12–14} Self-sustaining reentrant activity in inducible models consists of either AF-like rotor activity or atrial flutter (AFL)-like macroreentrant tachycardia (MT).

A recent proof of concept study demonstrated the feasibility and efficacy of ablating AF targets identified by computational simulations.¹⁵ In fact, a clinical trial based on that methodology is currently recruiting patients for randomization to computational modeling prior to ablation. The clinical workflow includes pre-ablation MRI acquisition and performance of simulations in patient-specific LA models. However, in prior studies conducted using the same methodology, 10–55% of pre-ablation models derived from patients with documented AF could not be paced into reentrant arrhythmia, that is, models were not inducible.^{7, 15–17} A non-inducible model from a patient with AF will not be able to guide an ablation. Thus, a significant priority is to better identify inducibility in models from AF patients, as this would allow us to extend the modeling approach to a broader range of patients with persistent AF. We propose to change the current clinical workflow by personalizing each patient’s electrophysiological properties, specifically ϴ.

To characterize the potential impact of ϴ personalization in simulations of the fibrotic LA, we investigated the effect of ϴ modulation on agreement between persistent AF models and clinically observed arrhythmias (i.e., phenotypes). We hypothesized improved agreement in the context of a wider but physiologically plausible range of potential ϴ values. We also investigated the relationship between model/phenotype agreement and risk factors for fibrosis and AF recurrence.

2. Methods

2.1. Study design

This is a retrospective, single-center cohort study conducted on consecutive persistent AF patients seen at the University of Washington (UW) between 2015 and 2019. IRB approval was secured (protocol #HSD 8763) as part of the UW Cardiac Arrhythmia Data Repository (CADRe). Study participants gave informed consent prior to participation and the investigation conformed to the principles outlined in the Declaration of Helsinki. Inclusion criteria were persistent AF at time of initial visit and availability of pre- and post-ablation LGE-MRI. Patients with prior AF ablation were excluded. Study data were collected and managed using the REDCap system, hosted at UW.^{18, 19}

2.2. Patient cohort

Patients had clinical assessment and catheter or surgical ablation at the UW AF program. All patients underwent pulmonary vein isolation, and some had additional substrate modification. Persistent AF status was determined at time of initial visit using standard Heart Rhythm Society consensus criteria.¹ Comorbidities and medications at the time of the initial visit were determined using electronic medical record review. LA volume index and surface area were measured on pre-ablation MRI. Patients underwent ablation and were followed longitudinally at UW with 7-day ambulatory ECG monitoring at 3, 6, and 12 months after ablation. Recurrence was defined by at least 30 seconds of documented atrial arrhythmia after a 90-day blanking period.¹ The type of recurrence was either AF or AFL in our patients. Patients without AF or AFL during at least 2 years of follow-up were deemed non-recurrent (NR). In some patients with recurrence, follow-up procedures were used to confirm mechanisms of recurrence. Loss-to-follow-up bias was limited because all patients completed at least 2 years of prespecified follow-up. There was no missing data for any patient in this analysis.

To address possible over-fitting or false positive models, in which models from patients without AF or recurrence would be made inducible, we used a series of control atria. We collected four LA models from patients with a range of ages, BMIs, and LA fibrosis percentages who were known not to have AF.

2.3. LGE-MRI acquisition

LGE-MRI of the LA was obtained on all participants to quantify fibrosis using previously described protocols.³ MRI acquisition characteristics are previously described and were identical for pre- and post-ablation imaging.¹⁶

2.4. 3D LA model construction, ϴ modulation, and AF simulations

Supplemental text is provided explaining the process of constructing post-ablation models, conducting AF simulations using a range of ϴ (from 0.8×ϴ to 1.2×ϴ, see Supplemental Figs. 1 and 2), and defining and assessing model/clinical phenotype agreement (Supplemental Fig. 2D).^{7, 8, 16, 20–31} For the baseline condition in normal tissue, the ϴ value in the longitudinal direction was 0.715 m/s; for the lowest and highest ϴ conditions, this was perturbed to 0.572 m/s and 0.858 m/s, respectively. In contrast, baseline ϴ in fibrotic tissue was 0.470 m/s and ranged from 0.376 m/s to 0.564 m/s. For each patient, one pre-ablation and one post-ablation model were created. In each model, pacing was performed at 15 different sites of common AF triggers, including the pulmonary veins, the appendage, the posterior wall, mitral valve, and the anterior wall.¹⁶ Each simulation was repeated for a total of 5 times (one each using 0.8, 0.9, 1.0, 1.1, and 1.2×ϴ).

2.5. Electroanatomic assessment of conduction velocity

Electroanatomic mapping was used to measure clinical ϴ in normal sinus rhythm during initial ablation when available. Further detail regarding this methodology is provided in the supplemental text.

2.6. Statistical analysis

Statistical analyses were performed using R [R Foundation for Statistical Computing, Vienna, Austria].³² Cohort descriptors and prevalence of comorbidities are presented as means with 95% confidence intervals for continuous variables with normal distributions, medians with interquartile ranges for non-normal distributions, and as proportions for binary variables.³³ We used the Shapiro-Wilk test to check for continuous variable distribution normality. We used box and whisker plots to represent the number of per model inducible simulations, rotors, and MTs. Outliers represented values 1.5×IQR above and below the 25^th and 75^th percentile values and were displayed as dots. We used McNemar’s test to compare the effect of ϴ modulation on model/phenotype agreement in the total cohort for each model/patient pair. We used Fisher’s exact test to compare the proportion of correct model/phenotype agreement in 2 instances: 1) subgroups of the population separated by presence or absence of baseline categorical risk factors (gender, hypertension, obstructive sleep apnea, and medication use) and 2) subgroups of the population separated by being above or below median values for baseline continuous risk factors (age, BMI, and percentage LA LGE). Fisher’s exact test is specifically used for each pair of “yes/no” clinical factors, and not between the ϴ themselves, because this allowed for the 2 comparisons listed above using each of ϴ. The risk factors chosen for analysis were those that had both a plausible link to clinical and simulation outcome and that were present in the cohort in sufficient proportion to allow for meaningful analysis. After testing for normality, we used the Wilcoxon signed rank test to compare the paired (pre- and post-ablation for the same patient) proportions of rotors to MTs.

3. Results

3.1. Baseline characteristics

Thirty-seven persAF patients (mean age 62.9 years, 56.8% male) were identified. The median time from pre-ablation MRI to ablation was 42 days (IQR 15–85) and the median time from ablation to post-ablation MRI was 249 days (IQR 181–346). The recurrence rate (AF or AFL) within 2 years of ablation in our cohort was 56.7% and the median time to recurrence was 144 days (IQR 116–242 days). Fifteen (41%) patients recurred with AF and 6 (16%) with AFL; 16 (43%) were NR. Sixteen patients underwent repeat ablation in the UW AF program. Table 1 shows demographic data for the study cohort.

Table 1:

Demographics of the Study Cohort

	Pre-Ablation	Post-Ablation Clinical Phenotypes
	Persistent AF	AF Recurrence	AFL Recurrence	No Recurrence
N (%)	37	15 (40.5%)	6 (16.2%)	16 (43.2)
Age (years)	62 (58–71)	66 (59–76)	67 (60–73)	59 (58–66)
BMI	31.8 (26.4–36.4)	31.8 (25.3–36.3)	32.5 (31.1–33.0)	31.54 (27–36.9)
Gender (male)	21 (56.8%)	8 (53.3%)	2 (33.3%)	11 (68.8%)
Prior DC cardioversions	1 (1–3)	1 (1–3)	3 (1–4)	1 (1–1)
Hypertension	21 (56.8%)	7 (46.7%)	3 (50%)	11 (68.8%)
Hyperlipidemia	12 (32.4%)	5 (33.3%)	1 (16.7%)	6 (37.5%)
Coronary artery disease	6 (16.2%)	3 (20%)	2 (33.3%)	1 (6.3%)
History of LVEF < 45%	9 (24.3%)	2 (13.3%)	2 (33%)	5 (31.3%)
OSA	11 (29.7%)	5 (33.3%)	1 (16.7%)	5 (31.3%)
Stroke/TIA	2 (5.4%)	1 (6.7%)	0 (0%)	1 (6.3%)
Diabetes	5 (13.5%)	2 (13.3%)	0 (0%)	3 (18.8%)
COPD	3 (8.1%)	2 (13.3%)	0 (0%)	1 (6.3%)
Tobacco use (Current/former)	7 (18.9%)	2 (13.3%)	1 (16.7%)	4 (25%)
Thyroid disease	4 (10.8%)	2 (13.3%)	2 (33.3%)	0 (0%)
Hypertrophic Cardiomyopathy	2 (5.4%)	2 (13.3%)	0 (0%)	0 (0%)
LA LGE (%)	17.9±2.5	18.8±4.7	17±7.2	17.3±3.8
LAVI (mL/m2)	61.1±21.5	61.6±17.2	67.1±13.6	54.2±29.6
LA surface area (cm2)	163±36.6	163±40.9	152±21.3	166±43.0
Prescribed amiodarone	10 (27.0%)	3 (20%)	2 (33.3%)	5 (31.3%)
Prescribed metoprolol	19 (51.3%)	8 (53.3%)	3 (50%)	8 (50%)
AF ablation type:
Cryo PVI only	13 (35.1%)	5 (33.3%)	3 (50%)	5 (31.3%)
Cryo PVI + RF SM	2 (5.4%)	1 (6.7%)	0 (0%)	1 (6.3%)
RF PVI only	5 (13.5%)	3 (20%)	0 (0%)	2 (12.5%)
RF PVI + SM	17 (45.9%)	6 (40%)	3 (50%)	8 (50%)
Surgical Cox-MAZE	1 (2.6%)	0 (0%)	1 (16.7%)	0 (0%)

Left atrial (LA) LGE was normally distributed and is reported as mean ± 95%CI; all other continuous variables are reported as median (IQR). For binary variables, proportions are reported. BMI, body mass index, Cryo, cryoballoon, COPD, chronic obstructive pulmonary disease, DC, direct current, LAVI, LA volume index, LGE, late gadolinium enhancement, LVEF, left ventricular ejection fraction OSAH, obstructive sleep apnea, PVI, pulmonary vein isolation, RF, radiofrequency, SM, substrate modification, TIA, transient ischemic attack.

3.2. Outcomes of simulations

In total, 5550 in silico pacing simulations (74 models × 15 pacing sites/model × 5 ϴ) were conducted in which 1078 (19.4%) simulations yielded either an MT or rotor. The total number of inducible simulations for pre- and post-ablation combined increased with each decrease in ϴ, and the proportion of rotors to MTs also increased slightly with decreasing ϴ (Fig. 1A). When simulation results for each ϴ value were considered separately for pre- and post-ablation models, pre-ablation model pacing induced more rotors than MTs, and post-ablation model pacing induced more MTs than rotors (Fig. 1B). The median proportion of model rotors (compared to MTs) was 0.70 pre-ablation and decreased to 0.25 post-ablation (p<0.001). The median number of inducible simulations per model increased with lower ϴ values in both pre- and post-ablation models (Fig. 1C). In pre-ablation models, this effect was driven largely by an increase in rotors (Fig. 1D), as there was little change in rotors in post-ablation models. In contrast, in post-ablation models there a less pronounced increase in MTs (Fig. 1E) accompanied by little change in MTs in pre-ablation models. The increase in post-ablation MTs drove the overall increase in inducible post-ablation simulations compared to pre-ablation. Taken together, these results indicate that 1) post-ablation models were significantly more likely to contain substrate for MTs than for rotors across the simulated ϴ range and 2) pre-ablation models were more likely to harbor more rotors at lower ϴ values while post-ablation models were more likely to harbor MTs at lower ϴ values.

Figure 1:

Panel A shows the total number of simulations in which pacing induced a rotor or MT in pre- and post-ablation models. Panel B shows the percentage of simulations that demonstrated rotors or MTs separated into pre- (left) and post-ablation (right). Panels C-E contain box-and-whisker plots showing the number (per model) of total inducible simulations, inducible rotors, and inducible MTs, respectively. All panels are stratified by ϴ, increasing from left to right. MT, macroreentrant tachycardia, ϴ, conduction velocity.

None of the four control LA models were inducible across the range of ϴ values tested. Supplemental Fig. 3 shows clinical characteristics, LA models, and activation maps in response to overdrive pacing for control patients. This highlights the absence of wavebreak and large-scale conduction block, which are prerequisites for reentry initiation.

3.3. Demonstration of model/phenotype agreement using ϴ modulation

Four examples of 3D LA models without phenotype agreement using baseline ϴ are shown in the left panels of Fig. 2. We used isochronal activation maps, shown in the right side of each panel, to demonstrate rotors and MTs induced when ϴ modulation was used.

Figure 2:

Panels A-D show four models that were not inducible using baseline ϴ that demonstrated rotors or MTs and agreement with clinical phenotype using the designated ϴ. The left side of each panel shows the non-inducible baseline model. The right side of each panel shows the corresponding activation maps that demonstrate simulation inducibility at each of the specified ϴ values. The top row (A-B) shows the effect of exacerbated ϴ (0.8 and 0.9×ϴ) and the bottom row (C-D) shows attenuated ϴ (1.1 and 1.2×ϴ). AF, atrial fibrillation, LAA, left atrial appendage, L/RIPV, left/right inferior pulmonary vein, L/RSPV, right/left superior pulmonary vein, MT, macroreentrant tachycardia, MV, mitral valve, ϴ, conduction velocity.

3.4. Evaluation of model/phenotype agreement

Fig. 3A summarizes model/phenotype agreement in the entire cohort, pre-ablation persistent AF models, and post-ablation models. When simulations were conducted using baseline ϴ values, model/phenotype agreement was observed in 21/37 (57%) pre-ablation models and 13/37 (35%) post-ablation models. Model/phenotype agreement in the total cohort using baseline ϴ was 46% (34/74 models). When any ϴ was allowed, total cohort model/phenotype agreement significantly improved to 65% (48/74 models, p=0.014). There was a trend toward improved model/phenotype agreement when any ϴ was allowed for pre-ablation models and post-ablation, AF-recurrent models. Most of the improved model/phenotype agreement in the overall cohort occurred in pre-ablation models and post-ablation, AF-recurrent models (Fig. 3B). No improvement in model/phenotype agreement was observed in AFL-recurrent patients using ϴ modulation. In one NR patient, model/phenotype agreement was observed using a non-baseline ϴ.

Figure 3:

Panel A shows model/phenotype agreement using baseline ϴ (left column in each set of columns) vs using any ϴ (right column). The Y-axis shows the number of models in agreement divided by the total number of models multiplied by 100. The total cohort is the left most set of columns and then each phenotype follows. Panel B shows the number of models for each phenotype and the total cohort that achieved model/phenotype agreement when different ϴ values were used, with increasing ϴ from left to right. Panel C shows model/phenotype agreement in the total cohort using each specific ϴ value, increasing from left to right. Panel D shows model/phenotype agreement using each specific ϴ value for each clinical phenotype. AF, atrial fibrillation, AFL, atrial flutter, ϴ, conduction velocity.

Combined pre- and post-ablation model/phenotype agreement increased as ϴ decreased (Fig. 3C). Improvement in model/phenotype agreement occurred in pre-ablation and AF-recurrent patients mainly when using decreased ϴ (Fig. 3D). Model/phenotype agreement was unchanged for AFL-recurrent patients across all ϴ and was unchanged for non-recurrent models except using 1.2×ϴ.

No constraints were placed on ϴ post-ablation relative to pre-ablation model/phenotype agreement. Therefore, a patient could have ϴ values for pre-ablation that achieved model/phenotype agreement that were not necessarily the same as the post-ablation ϴ values to achieved agreement. Conversely, there were 14 cases in which using the same (non-baseline) ϴ achieved model/phenotype agreement in both the pre- and post-ablation models. Therefore, a significant number of patients had a non-baseline ϴ value at which model/phenotype agreement was achieved that remained stable between two distinct time points. Most of these patients achieved pre- and post-ablation model/phenotype agreement using decreased ϴ values, although there were four models in which increased ϴ led to model/phenotype agreement. Most inducible models exhibited inducibility using multiple ϴ.

3.5. Relationship between pre-ablation clinical variables and model/phenotype agreement

Figs. 4A–D shows model/phenotype agreement stratified by pre-ablation clinical variables known to be associated with either AF or AF recurrence: BMI (<32 or ≥32), percentage fibrosis (<17.2% or ≥17.2%), hypertension, and obstructive sleep apnea, respectively. Model/phenotype agreement was significantly greater in patients with elevated BMI (using 0.8×ϴ) or elevated percentage LA fibrosis (using any ϴ) and trended toward greater in patients with hypertension or obstructive sleep apnea. There was no significant difference in model/phenotype agreement between older age (≥62 years) and younger age (<62 years) or between male and female gender (Supplemental Fig. 4).

Figure 4:

Panels A-D show model/phenotype agreement for patients separated by the presence of pre-ablation clinical characteristics: BMI, percentage LGE on pre-ablation MRI, OSA, and hypertension. Panels E-F show model/phenotype agreement for patients taking amiodarone and metoprolol, respectively. Each panel is stratified by ϴ, increasing from left to right. BMI, body mass index, LGE, late gadolinium enhancement, OSA, obstructive sleep apnea, ϴ, conduction velocity.

Figs. 4E–F show model/phenotype agreement in pre-ablation patients stratified by whether those patients were prescribed amiodarone and metoprolol prior to ablation, respectively. Both drugs are known to affect electrophysiological properties like ϴ. Patients taking either amiodarone or metoprolol, compared to patients not taking those medications, demonstrated improved model/phenotype agreement that was most apparent using 0.8×ϴ, although these did not meet statistical significance (e.g., using 0.8×ϴ, 78% of models for patients taking metoprolol agreed with clinical status compared to 56% for patients not taking metoprolol, p=0.17).

3.6. Conduction velocity measurements from electroanatomic mapping

Eight patients had clinical ϴ measurements available during normal sinus rhythm. On the posterior wall, the mean ϴ was 0.878 m/s (95% confidence interval: 0.138–1.61). These results are described in further detail in the Supplemental text. Supplemental Fig. 5 shows an illustrative example for the patient from whom we measured the lowest clinical ϴ in the posterior wall (0.462 m/s) of all eight patients. This patient’s model was not inducible for reentrant arrhythmia using baseline ϴ even though the patient was known to have persistent AF (i.e. no model/phenotype agreement at baseline). However, the model was inducible for multiple MTs using 0.8 and 0.9×ϴ, resulting in greatly improved model/phenotype agreement.

4. Discussion

Using state-of-the-art computational models created from pre- and post-ablation LGE-MRI scans of 37 persistent AF patients, we demonstrated that: 1) ϴ modulation improved model/phenotype agreement, most frequently with decreased ϴ, 2) MTs were significantly more common in post-ablation models and likely reflect a scar-driven arrhythmia substrate, and 3) baseline obesity or a high degree of LA fibrosis were associated with improved model/phenotype agreement. Our study has several strengths. We conducted AF simulations in models derived from a large, well characterized cohort of persAF patients who underwent ablation, had pre- and post-ablation LGE-MRI, and were followed closely for arrhythmia recurrence for two years. This was the largest study of paired pre- and post-ablation LGE-MRI-based computational simulations in which scar-geometry juxtaposition was used to create post-ablation models, as opposed to models built from post-ablation LGE-MRI alone. The extensive nature of this AF simulation study compared to many prior attempts^{7, 15–17} was thanks to the high degree of clinical characterization of patients in the CADRe database as well as the consistency in image acquisition. In prior studies, many patient images were not available for analysis³⁴ or did not have pre- and post-ablation imaging included in their studies.^{15, 16} Patients had detailed follow-up with serial ECGs, ambulatory monitoring, and symptom-triggered workup for 2 years after ablation, which was longer than in prior studies.^{7, 15–17, 34} We captured a mix of ablation approaches for AF, increasing the generalizability of our results. We corroborated patients’ recurrence rhythms on 12-lead ECGs, ambulatory monitoring, and using intracardiac mapping findings in patients who underwent repeat ablation procedure. We were able to differentiate mechanisms of recurrence in our study, while others included only AF recurrence or did not differentiate between types.^{17, 34} We included control studies in which no changes in inducibility were observed with changing ϴ. An important aspect of our work is linking in silico MTs to their clinical AFL-recurrent phenotype. Although most improvements in agreement occurred with decreased ϴ, we found four models with improved agreement using higher ϴ. Thus, our work shows that modulation rather than just ϴ reduction is key to improve agreement.

4.1. Comparing the effects of ϴ modulation in pre- and post-ablation models

In pre-ablation models, decreasing ϴ increased the number of rotors more than the number of MTs. In contrast, in post-ablation models, decreasing ϴ had a greater effect on raising the number of inducible MTs than rotors. The best results for improving model/phenotype agreement were in pre-ablation persAF or post-ablation, AF-recurrent patients. The effect of modulating ϴ on model/phenotype agreement in AFL-recurrent patients was not significant, indicating that those recurrences may be more related to simulated action potential duration, ablation lesion scar size, post-ablation residual (unablated) fibrosis, or emergent properties such as wavefront chirality (clockwise versus counterclockwise rotation).

It is known that decreased ϴ predisposes to reentry in animal models.³⁵ Indeed, we observed more inducible simulations (rotors, specifically) using decreased ϴ. Our study adds to the field by reporting this same trend in a set of computational models derived from a clinically well-characterized cohort, both pre- and post-ablation, with granular outcomes classification (no recurrence vs. AF recurrence vs. atrial flutter recurrence). Model personalization is still a relatively undefined area of research in the AF modeling space. We used a physiologically plausible range of values rather than a single value. Our work sets the stage for future studies that might better incorporate local heterogeneity in ϴ, as has been suggested by others using high density electroanatomic mapping of the LA.²⁵ Finally, while there were more rotors using lower ϴ, it was not a given that the models would exhibit more agreement with clinical outcomes, which was a novel finding.

4.2. Effect of baseline clinical characteristics on model/phenotype agreement

This study advances computational modeling of AF by addressing the relationship between in silico modeling parameters and clinical characteristics. Our results indicate that ϴ modulation improved model/phenotype agreement more in models derived from patients with comorbidities associated with AF-recurrence (especially elevated BMI and LA fibrosis). Putatively, the effects of pre-ablation clinical factors including medications on model/phenotype agreement are mediated by altered electrophysiological properties, including ϴ. Obesity, in a sheep experimental model, is associated with increased levels of pro-fibrotic TGF-β1³⁶ as well as increased epicardial adipose tissue, which is known to have paracrine effects such as altered local electrophysiology and increased fibroblast recruitment.^{37, 38} Specifically, in obese sheep with increased atrial adipose tissue deposition, increased TGF-β1, and increased burden of fibrosis, in vivo atrial ϴ was significantly reduced compared to non-obese controls.³⁶ In humans, as a result of factors like increased fibrosis, obesity is associated with decreased atrial ϴ when measured using intracardiac local activation times.²⁶

The improvement in model/phenotype agreement observed in patients with higher LGE is due to the increased number of rotors induced using lower ϴ values. Recent studies involving human patients presenting for AF ablation highlight the inverse relationship between burden of LA LGE and local/regional ϴ as estimated using intracardiac local activation times.^{25, 39} Therefore, low rates of model/phenotype agreement in low LA LGE may be explained by higher ϴ. Patients with extensive LA fibrosis are the most likely to have recurrence after ablation, and therefore are set to gain the most from pre-ablation identification of arrhythmogenic targets.

Patients taking amiodarone or metoprolol are likely to benefit from ϴ modulation pre-ablation. Larger studies of AF patients with paired LGE-MRI (such as in DECAAF II⁴⁰) should have the power necessary to detect statistically significant effects on ϴ modulation for clinical characteristics and medications. Such larger studies may also help elucidate the effects of post-ablation anti-arrhythmic medications on ideal modeling parameters.

In our study, we found that most of the model/phenotype agreement that was achieved using ϴ modulation involved lower rather than higher ϴ values, indicating that models with default parameters overestimate ϴ for many patients. In addition to identifying patients who are more likely to achieve better model/phenotype agreement using ϴ modulation, clinical factors may also identify patients in which baseline ϴ would be likely to overestimate the ideal model ϴ values. Conversely, ϴ values may not be over-estimated and instead the modulated models may be better at considering or capturing other pro-arrhythmic changes. We did note many cases in which the same non-baseline ϴ led to model/phenotype agreement pre- and post-ablation, but we also found a small number of cases in which a different ϴ pre- and post-ablation led to model phenotype agreement. More work is necessary to best determine the magnitude and effect of ϴ changes in pre- and post-ablation patients and their corresponding models.

4.3. Clinical perspective

There is growing interest in the use of AF simulations to identify novel targets for AF therapies,⁴ putting a premium on improving model/phenotype agreement pre- and post-ablation. AF models are already personalized for anatomy and fibrosis, and we show here that a strong next candidate for personalization is model fibrosis behavior. Ultimately, biomarkers of ϴ must be developed to adjust ϴ in models prior to ablation. There are several putative processes associated with atrial fibrillation and ϴ, including fibrosis, heart failure, aging, inflammation, hypertrophy, and ischemia. Plasma fibrosis biomarkers vary widely between patients¹⁰ and are associated with AF recurrence risk,⁹ making them an appealing option for a surrogate measure of pre-ablation ϴ. The current clinical workflow, as highlighted in the OPTIMA¹⁵ trial, includes pre-ablation imaging and simulation of AF. The clinical workflow change that our work suggests is that each patient’s ϴ should be personalized and applied to their model. Understanding factors that could suggest different fibrosis behavior, like antiarrhythmic medications, clinical characteristics, or fibrosis biomarkers, will be essential to broaden adoption of pre-ablation modeling to guide AF therapy delivery. Our work supports the concept that ablation can introduce substrate for MTs and must be balanced against the benefits of ablation of AF substrate. Finally, computational modeling of AF is a resource-intensive process,⁵ making it necessary to identify the patients (by certain clinical characteristics or biomarkers) most likely to benefit from pre-ablation simulations.

4.4. Limitations

A possible limitation of our work is that we identified patients’ recurrences as either AF or AFL. Our study dealt with this by allowing follow-up ablation procedure findings to overrule ambulatory monitoring, leading to the re-characterization (and probably more accurate characterization) of several patients’ clinical outcomes. For patients who recurred that received medical therapy or follow-up procedures outside of the UW program, recurrence mechanism descriptions relied solely on ECG or ambulatory monitoring. In addition, the simulation methodology used is not well-suited to represent focal arrhythmia compared to reentrant patterns. We note that model/phenotype agreement was worst in NR and was unchanged by modulation. This may have been due to insensitivity in the criteria we used to define model/phenotype agreement in NR patients but could have also been due to an inability to fully detect recurrences. Future work could consider alternative definitions or could increase ambulatory monitor or loop recorder use to better define recurrence.

Another possible limitation is the role that fibrosis progression in non-ablated tissue may play in disagreement between clinical and simulation outcomes. Progressive fibrosis affecting non-ablated tissue, which may occur between the two MRI scans, cannot be captured in post ablation scar imaging based on a higher hyper-enhancement threshold. This additional fibrosis may affect the results of computational simulations as well as clinical outcomes and therefore serve as a significant confounder. More work is necessary to delineate effects on ϴ from obesity and/or LA fibrosis. In our study, we did not include the right atrium, which could supply non-LA triggers of AF. Nonetheless, in one important study on right atrial triggers, LA models that were heavily fibrotic, as would be expected in persistent AF patients, were less likely to be dependent on right atrial triggers.⁴¹ Furthermore, most clinical ablation strategies for persistent AF are targeted to the LA and not the RA. Finally, the sample size, although large for this type of work, is still limited. Future work with a larger sample size of patients with pre- and post-ablation LGE-MRI, such as in the DECAAF-II⁴⁰ study, is necessary.

One possible confounder is that rotor activity itself may be more frequent at lower ϴ. For example, rotary wavefront activity and pivot points were shown to be associated with slower ϴ in patients with atrial fibrillation who underwent high density electroanatomic mapping.⁴² This introduces the risk for false positives. To mitigate this limitation, we demonstrated in four control cases of subjects without known AF that ϴ modulation did not introduce inducibility.

We were able to confirm that the range of simulated ϴ that we tested reflected similar values to that of clinical ϴ from electroanatomic mapping. Our study was not powered to compare the simulated ϴ to electroanatomic derived ϴ in all patients. Six of the eight models came from patients who already had model/phenotype agreement at baseline. The patient with the lowest clinical ϴ demonstrated improved model/phenotype agreement when lower ϴ were used in simulations. Many patients present for ablation in atrial fibrillation and cardioversion was not performed until after pulmonary vein isolation for confirmation of entrance and exit block, which typically occurs after the mapping portion of the procedure. Future studies will need a larger number of patients to address this limitation if cardioversion is not performed prior to mapping for clinical reasons. Our electroanatomic mapping-derived ϴ used point estimates rather than an average of all local ϴ that could be derived. Future studies should use emerging platforms or modules in electroanatomic mapping software to display ϴ histograms and/or LA maps of local ϴ.

4.5. Conclusions

In persistent AF patients, we showed a significant improvement in model/clinical agreement when a wide, physiologically plausible range of ϴ values were used, indicating that ϴ personalization could improve model usefulness. Our study lays the groundwork for linking computational model behavior to pre-ablation ϴ quantification, through detailed clinical characterization and/or measurement of plasma fibrosis biomarkers. This work improves our understanding of the relationship between in silico electrophysiological properties and persAF patients’ clinical characteristics.

Central Illustration:

Current protocols for simulating reentrant atrial arrhythmia use the same conduction velocity for all models despite known differences between patients. This study evaluated how model/phenotype agreement would be affected by allowing a wider range of potential conduction velocities in simulations. Models were created from pre- and post-ablation clinical imaging of the left atrium. Simulations assessed for inducibility rotor-driven reentry or macroreentrant tachycardia (MT). Patients were monitored for recurrence for 24 months with ambulatory ECG monitoring. Model/phenotype agreement was defined for each of the four clinical phenotypes. In the overall cohort, conduction velocity modulation improved model/phenotype agreement.

Clinical Perspective:

1. Clinical competency:

   1. Medical Knowledge – Pulmonary vein isolation is a cornerstone treatment of AF, but rates of success in persAF are significantly lower.
   2. Patient Care and Procedural Skills – Pre-AF ablation simulations of reentrant arrhythmia in persAF patients have the potential to improve ablation success. These protocols are being studied in clinical trials. This study seeks to enhance the representativeness of models, which will improve their adoption and bring them closer to the bedside.

2. Translational outlook: This study addresses two important aspects of translational medicine. First, this is a large study of computational modeling in a well characterized population with well characterized atrial models derived from cardiac MRI. These models are opportunities to test basic research findings regarding the representation of electrical properties in fibrotic left atria. Secondly, by improving the representativeness of models and capturing a wider range of physiologic conduction velocity, we make it more likely that pre-ablation simulations are utilized to improve outcomes in persAF patients. Showing the improvement in model agreement with clinical arrhythmias makes it more likely that ablation programs will acquire the necessary infrastructure (MRI, computational resources).

Abbreviations:

AF

atrial fibrillation

AFL

atrial flutter

BMI

body mass index

CADRe

cardiac arrhythmia database repository

ECG

electrocardiogram

LA

left atrial

LGE

late gadolinium enhancement

MRI

magnetic Resonance Imaging

MT

macroreentrant tachycardia

NR

non-recurrent

PersAF

persistent atrial fibrillation

TGF-β1

transforming growth factor-beta 1

UW

University of Washington

ϴ

conduction velocity