Transmural Heterogeneity in 3D Fibrosis Distribution Is Associated with Endo-Epicardial Electrical Dyssynchrony in Persistent Atrial Fibrillation

Abstract

Aim:

Despite the established role of endo-epicardial dissociation in the substrate of persistent atrial fibrillation (PsAF), the underlying mechanisms within the 3D atrial wall that lead to this phenomenon are not fully understood. Using personalized digital twin (DT) technology, we investigated the association between 3D transmurally heterogenous fibrosis architecture and endo-epicardial dissociation in PsAF.

Methods and Results:

LGE-MRI scans from 21 PsAF patients were used to generate bi-atrial DTs. Following virtual induction of reentry using programmed stimulation, a comparative analysis of the electrophysiological and structural properties of reentry-sustaining regions (RRs, n=91) against controls (non-RRs, n=73) was performed. RRs exhibited significantly greater endo-epicardial dissociation compared to non-RRs, evidenced by larger differences in median transmembrane potential (60.1 vs. 4.1 mV, p<0.001) and local activation times (55 vs. 14.9 ms, p<0.001). Lower dice similarity coefficient (DSC) was observed in RRs compared with non-RRs, median 0.731 [IQR 0.660–0.812] vs 0.880 [IQR 0.642–0.971], respectively; p=0.0009. In multivariable logistic regression, fibrosis percentage, transmembrane voltage difference, and DSC were independently associated with RR status, whereas LAT difference was not.

Conclusion:

Our findings reveal that 3D transmurally heterogeneous fibrosis architecture is associated with endo–epicardial electrical dissociation in PsAF, with increased dissociation observed in regions harboring reentrant circuits. Our findings advance the mechanistic understanding of PsAF arrhythmogenesis and provide a rationale for developing novel mapping and ablation strategies that account for the transmural complexity of the atrial substrate.

Introduction:

Atrial fibrillation (AF) is the most prevalent cardiac arrhythmia, imposing a significant healthcare burden, as it is associated with elevated risk of stroke, heart failure, and mortality.(Tsao et al., 2023) As the arrhythmia progresses from paroxysmal episodes to its persistent form (PsAF), the underlying mechanisms shift from ectopic triggers originating in pulmonary veins to self-sustaining activity rooted in the remodeled atrial tissue, or “substrate”.(Schotten et al., 2011) However, the specific properties within the diseased atrial substrate that sustain PsAF remain incompletely understood, representing a major challenge in the management of AF.(de Groot et al., 2025; Heijman et al., 2018) A deeper mechanistic understanding of the PsAF substrate and how it contributes to arrhythmogenesis is therefore crucial for developing more effective therapies that go beyond empirical treatment strategies.

A key electrophysiological phenomenon observed in patients with PsAF is the dissociation between endocardial and epicardial electrical activity. Foundational optical and electrical mapping studies in animal models,(Eckstein et al., 2011; Yamazaki et al., 2012) and subsequent high-resolution clinical mapping (De Groot et al., 2016; Maesen et al., 2022) have revealed significant differences in timing, activation patterns, and wavefront direction between the endo- and epicardial surfaces of the atria. This endo-epicardial dyssynchrony, which has been shown to manifest as epicardial wave breakthroughs,(De Groot et al., 2010) suggests the presence of intricate, three-dimensional conduction pathways that cannot be deduced from surface recordings alone. While these studies have identified endo-epicardial dissociation as a key manifestation of the PsAF substrate, whether the patient-specific 3D fibrotic architectures in the atria, a hallmark structural remodeling associated with PsAF, could underlie endo-epicardial dissociation and by what means has not been fully elucidated.

The presence of atrial fibrosis has been established as a critical component of the arrhythmogenic substrate in patients with PsAF. Visualized and quantified using late gadolinium enhancement (LGE) MRI, the extent of fibrosis is a strong predictor of AF recurrence after ablation and of adverse cardiovascular events. (Chelu et al., 2018; Siebermair et al., 2017) Evidence further indicates that the spatial distribution and characteristics of fibrosis contribute significantly to AF.(Verheule and Schotten, 2021) However, research has largely focused on the 2D distribution of fibrosis on the endocardial surface, as it is the surface most accessed during ablation procedures.

Personalized cardiac digital twins (DTs)—mechanistic computational models that represent the patient’s heart anatomy, structural remodeling distribution, and electrophysiological properties—offer a unique platform to provide such mechanistic insight into the role of the 3D fibrosis distribution in epi-endo- dissociation. By integrating patient-specific atrial structural data with biophysically detailed cell- level representations, DTs can simulate electrical wave propagation through the complex 3D tissue architecture of the disease-remodeled organ to explore arrhythmogenic mechanisms, an inquiry difficult to achieve through direct clinical or experimental studies. Our group and others have previously used this technology to provide a better understanding of AF mechanisms, revealing that re-entrant drivers in PsAF anchor to fibrotic boundary zones,(Zahid et al., 2016) elucidating how both preserved and newly emerged re-entrant drivers contribute to post-ablation recurrence of AF,(Ali et al., 2019) showing that rotational activity in AF is affected by remote lesions,(Sakata et al., 2024a) demonstrating that in PsAF the right atrial substrate has similar arrhythmogenic properties as the left,(Sakata et al., 2026b) and uncovering how some ablation lesions themselves can contribute to arrythmogenesis.(Yamamoto et al., 2026) The goal of this study is to leverage this validated and well-established personalized digital-twin technology to test the hypothesis that the transmural heterogeneity in the 3D distribution of fibrosis in the patient’s atria is associated with endo-epicardial electrical dissociation and increased substrate arrhythmogenicity in PsAF. This work provides a new facet in our understanding of the role of atrial fibrosis in arrhythmogenesis and the causes for endo-epicardial dissociation in PsAF.

Methods:

Ethical Approval

The study protocol was approved by the Johns Hopkins Institutional Review Board (IRB00183327) and was conducted in accordance with the Declaration of Helsinki, except for registration in a database. All patients provided written informed consent prior to enrollment in the study.

Study Overview

In this study, we utilized personalized bi-atrial DTs to examine associations between three-dimensional fibrosis architecture, endo-epicardial electrical dissociation, and regions capable of sustaining atrial reentry (Figure 1). LGE MRI scans were acquired for a cohort of 21 PsAF patients. Each patient’s imaging data was then used to construct a personalized, bi-atrial DT with patient-specific anatomy and fibrosis distribution, as visualized on the LGE. Next, a virtual electrophysiological (EP) study, involving programmed electrical stimulation, was conducted in each atrial DT to induce and locate regions that sustain reentrant activity. Finally, we compared the EP and structural properties of reentry-sustaining regions to those of control regions to identify structural and electrophysiological features associated with endo-epicardial electrical dissociation and the presence of RRs in the PsAF fibrotic substrate.

Figure 1:

Study overview (representative examples, n=3; protocol was applied to N=21 patients).

Study Cohort

The study included 21 patients with symptomatic, drug-refractory persistent AF who presented to Johns Hopkins Hospital, Baltimore, USA (data collected as part of Sakata et al., 2024a).

LGE MRI Acquisition

To re-construct the patient anatomy and characterize the atrial fibrotic substrate, cardiac MRI studies were performed. LGE MRI scans were acquired in sinus rhythm using 1.5 Tesla MRI scanners (Aera, Siemens Healthineers, Erlangen, Germany and Ingenia, Philips, Best, the Netherlands) and 3 Tesla MRI scanner (Achieva, Philips, Best, the Netherlands). To ensure stable sinus rhythm and optimize imaging quality, anti-arrhythmic drugs were administered to the patients, and, when needed, direct current cardioversion was applied a few hours before MRI study. The LGE-MRI scan was initiated 20–25 minutes following gadolinium-based contrast agent administration using a fat-saturated 3D inversion recovery-prepared fast spoiled gradient recalled echo sequence with respiratory navigation and ECG-gating. The typical scan parameters for the LGE scan were transverse imaging volume covering left and right atria, in-plane resolution 1.25×1.25 mm, slice thickness 2.5 mm, and reconstructed voxel size of 0.625×0.625×1.25 mm. The 3D LGE scan imaging data were acquired at end of diastole, with the inversion time for the scan determined by a TI-scout of left ventricle. Detailed description of the LGE-MRI protocol was provided in a previous publication.(Sakata et al., 2024a)

Construction of Personalized bi-atrial DTs

The computational approach to generate patient-specific bi-atrial DTs, encompassing both the left and right atria, has been comprehensively described in our previous studies.(Ali et al., 2019; Boyle et al., 2019, 2018; Hakim et al., 2018; Sakata et al., 2024a, 2024b; Zahid et al., 2016) A brief description is provided below (Figure 1B). The LGE-MRI scans were semi-automatically segmented using LASSnet, a convolutional deep neural network for bi-atrial segmentation.(Lefebvre et al., 2023) Subsequently, manual quality control was conducted using the open-source image-processing software ITK-SNAP to identify and correct any potential segmentation errors. Next, fibrotic tissue was distinguished from non-fibrotic tissue using the personalized image intensity ratio (PIIR) method described in our previous publications.(Sakata et al., 2026b, 2024a) The PIIR method calculated the intensity threshold for atrial fibrosis using the mean values of the atrial blood pool and a reference fibrous tissue; this approach was specifically utilized for its validated reproducibility across repeated LGE-MRI acquisitions, mitigating the threshold sensitivity observed in standard intensity ratio approaches. A 3D tetrahedral mesh with a target edge length of 300 μm was generated from the segmented and carved LGE-MRI images, incorporating the patient-specific fibrosis distribution, using Mimics software (Materialise). (Sakata et al., 2025, 2024a). The resulting meshes had an edge-count-weighted mean edge length of 358.51 ± 69.27 μm. Sensitivity to mesh resolution has been evaluated in prior work from our group.(Boyle et al., 2021) Fiber orientations were computed from human atlas diffusion tensor-MRI data(Pashakhanloo et al., 2016) and registered onto patient-specific meshes using a custom 3D mapping software and the Universal Atrial Coordinates 2 method.(Ali et al., 2021; Roney et al., 2019) EP properties were assigned to the non-fibrotic tissue and fibrotic remodeled substrate in PsAF patients’ DTs in accordance with our previous publications.(Ali et al., 2019; Boyle et al., 2019, 2018; Hakim et al., 2018; Sakata et al., 2024a, 2024b; Zahid et al., 2016) In brief, a human chronic atrial fibrillation (AF) action potential model was implemented for non-fibrotic tissue, while fibrotic regions were simulated by reducing maximal I_K1, I_CaL, and I_Na channel conductances by 50%, 50% and 40% respectively to represent pro-fibrotic signaling. These distinct ionic properties and assigned anisotropy ratios (non-fibrotic=5:1 vs fibrotic=8:1) resulted in different longitudinal conduction velocities for non-fibrotic (43.39 cm/s) and fibrotic tissues (20 cm/s), when tested on planar propagation in a tissue slab.(Hakim et al., 2018; Konings et al., 1994) Conduction robustness in our meshes was confirmed as shown in Figure 2. Finally, all simulations were performed on a parallel computing system using openCARP, a software freely available for academic use.(Plank et al., 2021)

Figure 2:

(A) An example illustrating our mesh discretization across the different tissue types, showing how the adaptive meshing algorithm refines the spatial resolution in the fibrotic regions (the values shown are in μm; yellow is a healthy region and red is fibrosis). (B) Simulations were performed on a 1×0.5×0.5 cm block at the resolution (231 μm) of fibrosis and with the corresponding lowest conductivity values to show that the wavefront propagates smoothly between adjacent nodes in one time step, with no evidence of spurious conduction block, confirming that the spatial and temporal discretization is sufficient for accurate propagation.

Assessment of Arrhythmogenicity in the DT Atrial Substrate

Virtual EP stimulation was performed in the DTs to induce substrate-based arrhythmia. Using an incremental burst pacing protocol, we paced each DT from 40 randomly selected bi-atrial sites—representing potential ectopic trigger locations within the arrhythmogenic substrate.(Boyle et al., 2021, 2019; Sakata et al., 2024a, 2024b) The detailed pacing protocol is presented in our previous publications.(Sakata et al., 2026b, 2025)

Reentry-sustaining regions (RRs) were defined in a multi-step process to capture the complex 3D intramural dynamics (Figure 1C). First, simulated transmembrane voltage (Vm) dynamics were visualized using the open-source visualization tool Meshalyzer (https://git.opencarp.org/openCARP/meshalyzer/), where depolarizing wavefronts were displayed using a Vm range of −40 to 0 mV at a temporal resolution of 20 ms. Next, the surface manifestation of sustained reentrant activity, i.e., a rotor-core/phase-singularity-like trajectory, was visually identified directly from the circulating Vm wavefront pattern. This process was double-checked under expert supervision. A reentrant activity was considered sustained if it persisted for at least 5 seconds and if its rotor-core trajectory remained repeatedly inscribed on at least one atrial surface (endocardial or epicardial) over consecutive rotations. In contrast, transient wavebreaks, wavefront collisions, breakthrough patterns, passive activation from adjacent sources, or non-recurrent conduction block were excluded. After marking the center of the PS trajectory, we performed a 3D volumetric analysis of the underlying wavefront dynamics using Paraview software. The depolarization wavefront was traced inside the atrial wall over consecutive cycles to delineate the complete reentrant pathway in 3D. An RR was defined as the minimal 3D bounding cylinder across the atrial wall that encompasses the complete, multi-cycle reentrant pathway. RR dimensions were determined such that if the core exhibited limited spatial meander while remaining within the same anatomical region, the entire meandering trajectory was included in the RR. All RR identifications were based on the same predefined criteria. Importantly, for reproducibility and downstream analyses, the rotor core location, cylinder radius, corresponding time window, and all mesh points with their data contained within the bounding cylinder were saved for each identified RR. Independent expert review yielded complete concordance for RR identification, and no cases required adjudication.

For comparative analysis, we defined non-RR control regions in each patient DT. These non-reentrant regions were sampled from predefined anatomical locations (LA: posterior, anterior, roof; RA: posterior, anterior, lateral). Any location containing a previously identified RR was excluded from the control selection process.(Sakata et al., 2026a) To assess whether differences in analyzed region size could confound comparisons between RRs and non-RRs, the diameter of each RR and non-RR was quantified.

Assessment of the Endo-epicardial Asynchrony

In each RR and non-RR, the presence of transmural electrical dissociation was assessed by calculating the transmembrane voltage differences and local activation time (LAT) differences between endocardial and epicardial layers as obtained from DT simulations. The transmural voltage difference was determined by identifying the frame with the maximum voltage difference during one complete rotor rotation and then computing the median point-to-point transmembrane voltage difference between the endocardial and epicardial surfaces. Similarly, the LAT difference was calculated by isolating the temporal window of rotor formation and computing the median point-to-point LAT difference between the two surfaces. Qualitative assessment of RRs was performed to classify them as asynchronous or synchronous, with classifications validated by multiple operators.

To assess whether endo-epicardial electrical differences were also present outside reentrant activity, we performed an additional stable-pacing analysis in the same anatomical RR and non-RR regions. Stable pacing was delivered from high right atrium at a cycle length of 600 ms, and the final stabilized paced beat was analyzed.

Additionally, the extent of atrial fibrosis on each surface was characterized by calculating the percentage of fibrosis in endocardial and epicardial layers. Furthermore, dice similarity coefficient (DSC) was calculated to assess spatial overlap between endocardial and epicardial fibrosis distributions. Within each RR, conduction pathways were analyzed and classified based on their transmural extent (semi-transmural or fully transmural), and their average width was determined.

Statistical Analysis

All data are reported as median with interquartile range (IQR), unless otherwise specified. Statistical comparisons were conducted using the Wilcoxon rank-sum test. A p-value < 0.05 was considered statistically significant. Univariable and multivariable logistic regression analyses were performed to evaluate associations between RRs and fibrosis percentage, transmembrane voltage difference, LAT difference, and DSC. Odds ratios are reported with 95% confidence intervals. For the matched case-control analysis, non-RR control regions were matched to RRs by atrial chamber and mean fibrosis density, and conditional logistic regression was used to evaluate the association between DSC and the presence of RRs within matched pairs. All statistical analysis was performed using R Statistical Software (v4.2.2; R Core Team, 2021, R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria).

Generative AI Disclosure

ChatGPT (OpenAI; GPT-5.x, exact version not recorded) and Gemini (Google; Gemini 3.x, exact version not recorded) were used for correcting grammar and sentence structure during preparation of the manuscript. Claude (Anthropic; Claude 4.x, exact version not recorded) was used to assist with coding. All AI-assisted text and code were thoroughly reviewed, edited, and approved by the authors before inclusion in the manuscript.

Results:

Cohort characteristics and arrhythmogenicity assessment

Patient characteristics are summarized in Table 1. The overall atrial fibrosis burden was 14.4% (IQR: 10.9–21.0%), with a greater fibrosis burden observed in the left atrium (LA), and most patients presented with enlarged atrial volumes, consistent with structural remodeling in PsAF. We ran more than 1600 induction simulations across the cohort and identified and analyzed 91 RRs and 73 non-RRs. These arrhythmogenic and non-arrhythmogenic sites were comparably distributed between the left and right atria (LA:38 RRs, 41 non-RRs; RA: 53 RRs, 32 non-RRs), underscoring the value of bi-atrial modeling for understanding PsAF mechanisms. The median diameter was similar between the chosen RRs and non-RRs, with RRs having a median diameter of 25.54 mm (IQR 22.07–29.19 mm) compared with 25.56 mm (IQR 23.01–28.05 mm) in non-RRs; this difference was not statistically significant by Wilcoxon rank-sum test (p = 0.8413).

Table 1:

Cohort Characteristics (n=21 patients). Data are reported as median [IQR: Interquartile range] for continuous variables and counts for categorical variables.

Characteristic	Value (n=21)
Age (years)	67 [58–72]
Female | Male	4 | 17
BMI	30.8 [27.6–33.9]
Ejection Fraction	58 [48–60.5]
CHA2DS2-VASc	3 [1.5–3.5]
Fibrosis Burden (from MRI)	14.4% [10.9–21.0%]
Patient with history of previous ablation | Patients without	5 | 16

Majority of RRs exhibit significant endo-epicardial electrical dissociation

A qualitative analysis of the transmembrane potential in the DTs revealed that RRs predominantly exhibited a marked dissociation in EP activity between the endocardial and epicardial surfaces (68%), with wavefronts propagating asynchronously through the atrial wall (Figure 3A). In contrast, all non-RRs and some RRs (32%) demonstrated near-synchronous activation on the endo- and epicardium (Figure 3B, 3C). Another illustration of a representative asynchronous RR case is shown in Figure 4A.

Figure 3:

Transmembrane potentials on the endocardium and epicardium in regions capable of sustaining reentrant activities (RRs) vs non-RRs (representative examples, n=3; protocol was applied to N=21 patients).

Figure 4:

(A) Reentrant activity on endo- and epicardial surfaces in an example RR in a patient’s DT (representative example, n=1), (B) Differences between endo- and epicardial median transmembrane potential in RRs (n=91 in N=21 patients) vs non-RRs (n=73 in N=21 patients) among all DTs, (C) Differences in endo- and epicardial median late activation times (LATs) in RRs (n=91 in N=21 patients) vs non-RRs (n=73 in N=21 patients). Red: Endocardial Wavefront; Green: Epicardial Wavefront; Blue: Median transmembrane voltage difference; Pink: Late activation time difference; Box plots represent the median (horizontal line), and interquartile range (box boundaries). Statistical significance was determined using the Wilcoxon rank-sum test. ***: P<0.001.

Next, the electrical dissociation in RRs was quantitatively analyzed. The median difference in transmembrane potential between endocardium and epicardium was significantly greater in RRs (60.1 mV [IQR: 55.4–63.9 mV]) compared to non-RRs (4.1 mV [IQR: 1.3–14.7 mV]) (p<0.001), as illustrated in Figure 4B. This dissociation was further reflected in activation times; RRs exhibited a significantly higher median endo-epicardial LAT difference of 55ms [IQR: 31.4 – 101ms] versus 14.9 ms [IQR: 3.6 – 41.3 ms] in non-RRs (p<0.001) (Figure 4C). Notably, the LAT difference values in RRs closely aligned with the transmural activation delay (67 ± 31 ms) reported in the ex vivo human heart study by Hansen et al.(Hansen et al., 2015)

To determine whether endo-epicardial electrical differences were present outside reentrant activity, we analyzed stable-pacing simulations in the same anatomical RR and non-RR regions for which stable-paced data were available. During stable pacing, LAT differences were small in both groups and were not significantly different between RRs and non-RRs: 3.82 ms [IQR: 2.92 – 5.46 ms] in RRs versus 3.32 ms [IQR: 2.41 – 5.01 ms] in non-RRs (p=0.085). In pointwise analysis of transmembrane voltage, RRs showed higher endo-epi voltage differences than non-RRs during stable pacing: 44.77 mV [IQR: 5.99 −69.30 mV] versus 10.88 mV [IQR: 3.87 – 52.09 mV], respectively (p=0.238).

Across analyzed regions, lower endo-epicardial fibrosis overlap was associated with greater LAT difference, as reflected by an inverse correlation between DSC and LAT difference (R²=0.57, p<0.001). We then performed univariable and multivariable logistic regression analyses to evaluate associations with RR. In univariable analyses, fibrosis percentage, transmembrane voltage difference, and DSC were each associated with RR. In the multivariable model, fibrosis percentage, transmembrane voltage difference, and DSC remained independently associated with RR, whereas LAT difference did not. Each 10% increase in fibrosis percentage was associated with higher odds of RR status (OR 1.86, 95% CI 1.14–3.03, p=0.013), as was each 10-mV increase in transmembrane voltage difference (OR 2.59, 95% CI 2.01–3.32, p<0.001).

Because DSC showed an adjusted association with RR status in the full multivariable model, we further evaluated whether transmural fibrosis overlap distinguished reentrant from non-reentrant regions after closer control for atrial chamber and local fibrosis density. For this purpose, we performed a matched case-control analysis. Non-reentrant control regions were matched to reentrant regions by atrial chamber and mean fibrosis density. Conditional logistic regression was then used to evaluate the association between DSC and reentrant-region status within matched pairs. In chamber- and fibrosis-matched regions using a ≤5 percentage-point fibrosis caliper, DSC was not independently associated with reentrant status, OR 1.23 per 0.1 increase in DSC, 95% CI 0.90–1.67, p=0.186. However, because matching reduced the analytic sample to 40 matched pairs, these results should be interpreted cautiously and considered hypothesis-generating rather than conclusive.

Transmural fibrosis characteristics in RRs and non-RR control regions

We next compared the fibrotic characteristics of all RRs and non-RRs. We found that RRs were characterized by significantly higher percentage of fibrotic tissue on both the endocardial surface (median 25.6% vs. 6.9% in non-RRs, p<0.001) and the epicardial surface (median 21.9% vs. 7.5% in non-RRs, p<0.001) (Figure 5A). Lower DSC was observed in RRs compared with non-RRs, median 0.731 [IQR 0.660–0.812] vs 0.880 [IQR 0.642–0.971], respectively; Wilcoxon rank-sum test, p<0.001. (Figure 5B). Qualitatively, both asynchronous and synchronous RRs were observed (Figure 5C). The mismatch of fibrosis distribution on the epi- and endocardial surfaces in asynchronous RRs provide the substrate for the observed functional EP dissociation, allowing wavefronts to propagate nearly independently on each surface. In contrast, the synchronous RRs were associated with more aligned, transmurally-continuous fibrosis. This structure acts as a single, cohesive obstacle, forcing the wavefront to propagate synchronously around it on both the endocardium and epicardium. These results demonstrate that while a substantial degree of fibrosis creates a general propensity to reentry, it is the specific 3D architecture of that fibrosis that dictates the type of reentrant activity manifested.

Figure 5:

(A) Percentage of epicardial and endocardial fibrotic substrate in RRs (n=91 in N=21 patients) vs non-RRs (n=73 in N=21 patients), (B) Dice similarity coefficient between endo- and epicardial fibrotic substrate RRs (n=91 in N=21 patients) vs non-RRs (n=73 in N=21 patients), (C) Representative examples of fibrotic substrate in RRs with endo-epicardial asynchrony and synchrony. Orange: Percentage endocardial fibrosis; Green: Percentage epicardial fibrosis; Purple: dice similarity coefficient. Black and white pattern: Epicardial fibrosis; Red: Endocardial fibrosis. Box plots represent the median (horizontal line), and interquartile range (box boundaries). Statistical significance was determined using the Wilcoxon rank-sum test. **: P<0.01; ***: P<0.001.

Figure 6 illustrates a representative wavefront dynamics in asynchronous RRs. The wavefront would alternate between fully transmural tissue and semi-transmural channels, rendering it intermittently invisible on either the endocardial or epicardial layers for certain periods, leading to endo-epi electrical dyssynchrony. Figure 7 illustrates the locations of all identified RRs and their 3D fibrotic architecture in the LA of one patient’s DT. Further, our analysis revealed that RRs were comprised of semi-transmural or fully transmural conduction pathways (Figure 8). Across all RRs we identified 306 conduction pathways; 71% of them were semi-transmural, while 29% were fully transmural, with a median of 2 semi-transmural pathways per RR (IQR: 1–4), compared to a median of 1 fully transmural pathway per RR (IQR: 0–2). The bulk of the channels were narrow, with a median width of 2.2 mm (IQR: 1.5–3.1 mm).

Figure 6:

Sequence of intramural wave dynamics that results in endo-epicardial dyssynchrony within the PsAF fibrotic substrate (representative example, n=1; protocol was applied to N=21 patients). Fibrosis is shown in gray.

Figure 7:

An example of an LA for a patient-specific DT with RR regions in red. The 3D fibrosis pattern in each isolated region is shown with the number of narrow channels, their width and their types: semi-transmural endo/epi or fully transmural channels. Representative example, n=1; protocol was applied to N=21 patients.

Figure 8:

Conduction channels in regions capable of sustaining reentrant activities (RRs; n= 91 in N=21 patients), with their types, distribution, average count per RR, and their widths. Green: Semi-Transmural Channel; Yellow: Fully Transmural Channel; Black and White Pattern: Epicardium; Gray: Endocardium. Box plots represent the median (horizontal line), mean (‘×’ marker), and interquartile range (box boundaries). Statistical significance was determined using the Wilcoxon rank-sum test. ***: P<0.001.

These observations allowed us to classify RRs into two main types (Figure 9): Type-I RRs, which consist of semi-transmural narrow conduction channels with or without fully transmural channels (87%, 79/91 RRs), and Type-II RR, which consists of only fully transmural narrow conduction channels (13%, 12/91 RRs). Within the dominant Type-I category of RRs, we observed two primary morphological patterns: 1) RRs that featured semi-transmural channels located exclusively on either the endocardial or epicardial surface, 2) RRs that were characterized by both endocardial and epicardial semi-transmural channels. The prevalence of Type-I RRs demonstrates that a substrate assessment based solely on endocardial mapping may be insufficient to characterize atrial arrhythmogenesis.

Figure 9:

Two main types of arrhythmogenic substrate: (A) Type-I, semi-transmural narrow conduction channels on one side (endo or epi) or both endo-epi sides, with or without fully transmural channels; (B) Type-II, exclusively fully transmural narrow conduction channels. Representative examples, n=3; protocol was applied to N=21 patients.

Discussion:

In this study, we leveraged personalized bi-atrial DTs to identify an association between heterogeneous 3D atrial fibrosis distribution and endo-epicardial dyssynchrony in PsAF. While the presence of atrial fibrosis is well understood as an important contributor to arrhythmogenesis in patients with PsAF, fibrosis has typically been assessed by its total amount (i.e. fibrosis burden),(Marrouche et al., 2014) or as an endocardial distribution,(Sim et al., 2019) and the impact of fibrosis 3D heterogeneity has not been fully appreciated as part of our understanding of the PsAF substrate. Here, we demonstrated that 3D re-entrant circuits persist in a substrate of highly heterogeneous and predominantly non-transmural fibrosis. We found that the majority of the atrial regions where the reentrant drivers persist are characterized with different distributions of fibrosis on endo- and epicardial surfaces and increased dyssynchrony of electrical activity between these surfaces. Additionally, the 3D distribution of fibrosis in RRs was found to often incorporate macro-channels that are only partially transmural. The additional stable-pacing analysis showed small and comparable LAT differences in RRs and non-RRs, despite higher pointwise voltage differences in RRs, suggesting that pronounced activation-time dyssynchrony is most evident during reentrant wave propagation rather than during stable-paced activation.

In the full multivariable model, fibrosis percentage, transmembrane voltage difference, and DSC were independently associated with RR status, whereas LAT difference was not. However, the direction of the DSC association differed between univariable and multivariable analyses and DSC was not independently associated with RR status in the chamber- and fibrosis-matched analysis. Thus, DSC should be interpreted as part of a broader structural-electrophysiological association with RR status rather than as a stand-alone determinant of reentry. Notably, these findings were observed in both the left and right atria, emphasizing the fact that PsAF is not an LA disease, despite current treatment strategies being primarily focused on the LA.

While the dissociation in electrical activity between endo- and epicardial activity during AF is well documented,(De Groot et al., 2016; Eckstein et al., 2011) starting with early canine studies that first hypothesized an anatomical basis for discordant endo-epicardial activation,(Schuessler et al., 1993; Zaĭtsev, 1989) the current understanding of the mechanisms underlying it is based on the work by Allessie, Schotten, and colleagues in goat animal models. In the paradigm based on the goat studies, AF was sustained by transmural chaotic wave breakthroughs, occurring because the epi- and endocardial atrial layers were found to not be well connected electrically, with structural discontinuities existing in the mid-myocardial layers of the atrial wall. When an electrical wave propagating on one surface encounters a break in the mid-myocardial layers, it emerges on the other surface as a breakthrough.(De Groot et al., 2016, 2010; Eckstein et al., 2013, 2011) This contrasts with other mechanistic paradigms that attribute AF maintenance to stable, localized re-entrant drivers anchored to structural fibrotic substrate.(Boyle et al., 2018; Hakim et al., 2018; Hansen et al., 2015; Kulathilaka et al., 2025; Sakata et al., 2024a; Zahid et al., 2016) Our study suggests that these two mechanistic frameworks are not mutually exclusive but may fundamentally represent the same phenomenon viewed through the lens of 3D substrate architecture.

Other animal studies, such as the one by Yamazaki et al. observed drifting rotors seen from endo or epicardium using ex vivo optical mapping in sheep atria. These were attributed to the intramural heterogeneity in atrial wall thickness and stretch, which were identified as key drivers in AF maintenance.(Yamazaki et al., 2012) Hansen et al.(Hansen et al., 2015) conducted an ex vivo human atrial study integrating high-resolution 3D structural imaging with simultaneous endo-epicardial mapping, confirming the presence of significant differences between endo and epicardial activation during sustained AF. The study attributed the presence of stable micro reentrant circuits inside the wall of atria as the potential mechanism that leads to endo epicardial dissociation in electrical activity. Interestingly, the in-silico LAT endo-epicardial discordance observed in our human-derived DT models closely matched the transmural activation delay reported in the ex vivo human heart study by Hansen et al.(Hansen et al., 2015) While these studies established the existence of endo-epicardial dissociation during AF and hypothesized that structural discontinuities underlie this phenomenon, our study demonstrated an association between transmural fibrotic heterogeneity and endo-epicardial electrical dyssynchrony in human atria. This finding helps unite different views of endo-epi dissociation, offering a more complete understanding of the complex substrate of PsAF.

The 3D atrial fibrosis distribution in patients with PsAF has remained overlooked because of the clinical intra-procedural view of fibrosis as an endocardial distribution, but also because majority of the patient-specific atrial DTs ignore the thickness of the atrial wall. While there have been numerous DT studies that have addressed the role of fibrosis distribution in AF,(Bayer et al., 2016; Labarthe et al., 2014; Macheret et al., 2023; Martínez Díaz et al., 2024; Roney et al., 2022, 2020) these studies have utilized atrial DTs that are geometrically only surface shells, thus assuming that the fibrosis is transmurally homogenous. Our atrial DTs encompass 3D bi-atrial geometries that capture the complex transmurally heterogenous fibrosis within the atrial wall, allowing for exploration of the arrhythmogenic mechanisms stemming from the 3D fibrosis distribution in PsAF.

Our findings suggest that the 3D structural heterogeneity of fibrotic tissue may warrant a more sophisticated targeting strategy during clinical ablation of PsAF. The presence of semi-transmural conduction pathways could serve as a novel target for substrate-based ablation, potentially leading to improved clinical outcomes over current empirical approaches. Furthermore, our findings demonstrated that significant number of semi-transmural channels associated with reentrant activity may be epicardial. This underscores a fundamental limitation of the current clinical standard of using endocardial bipolar mapping to assess the atrial substrate. Recent works, including findings by our own group, have begun to utilize unipolar electrical mapping to reveal deeper intra-mural and epicardial substrates.(Ali et al., 2025; Qi et al., 2025; Yavin et al., 2023) A hybrid mapping assessment incorporating unipolar voltage signals may be useful to identify the potential 3D PsAF substrate. Such an approach would also ensure no arrhythmogenic epicardial conduction pathways are left behind, a common failure of transmurally incomplete ablation lesions. The insights obtained here could thus pave the way for new strategies to achieve durable freedom from PsAF.

Limitations:

Our study included 21 patient-specific DTs (with >1600 simulations run on high-performance computing clusters), a sample size that may affect generalizability. Furthermore, although the median width of our semi-transmural channels was larger than the voxel size of the MRI acquisitions, our findings remain limited by the imaging resolution of our MRI acquisitions and do not capture the contribution of smaller-scale attributes to the fibrosis distribution. Another limitation is the selection of non-RRs might have introduced bias; however, 73 locations were systematically selected from typical regions of the atria across all patient cases, which can be considered sufficient for comparison with the 91 identified RRs.(Sakata et al., 2026a) Also, our DTs have constant wall thickness and although our protocol is consistent with Varela et al., the native atrial wall exhibits variable thickness.(Varela et al., 2017) Most recently, Kulathilaka et al. showed that wall thickness was associated with reentry distribution.(Kulathilaka et al., 2025) Consequently, our analysis does not account for the potential contribution of thickness heterogeneity to the observed endo-epicardial dissociation. In addition, the present analyses identify associations among transmural fibrosis distribution, endo-epicardial electrical dyssynchrony, and RR status, but do not establish causal direction. In particular, because dyssynchrony was primarily quantified during reentrant activity, these analyses cannot determine whether endo-epicardial dyssynchrony precedes rotor formation, facilitates reentry maintenance, or occurs as a consequence of sustained reentry.

Conclusion:

This study demonstrates an association between the 3D transmural fibrosis distribution and the endo-epicardial electrical dissociation in PsAF by integrating patient-specific 3D fibrotic architecture into fully volumetric bi-atrial DTs. We demonstrated that 3D atrial re-entrant circuits that sustain PsAF persist in a substrate of highly heterogeneous and predominantly non-transmural fibrosis and are associated with endo-epicardial electrical dyssynchrony. The 3D distribution of fibrosis in RRs was found to often incorporate macro-channels that are partially transmural, through which the reentrant waves propagate. Our work provides an additional facet in the mechanistic understanding of arrhythmogenesis in PsAF. It also provides a rationale for developing new mapping and ablation strategies in PsAF patients that are based on a more precise understanding of the transmurality of the atrial substrate.

Table 2:

Univariable and multivariable logistic regression analysis for RR

Model	Predictor	Odds ratio	95% CI	p-value
Univariable	Fibrosis percentage, per 10%	1.60	1.28–2.01	<0.001
Univariable	transmembrane voltage difference, per 10 mV	2.39	1.93–2.97	<0.001
Univariable	LAT, per 10 ms	1.02	0.99–1.05	0.174
Univariable	DSC, per 0.1increase	0.82	0.68–0.98	0.026
Multivariable	Fibrosis percentage, per 10%	1.86	1.14–3.03	0.013
Multivariable	transmembrane voltage difference, per 10 mV	2.59	2.01–3.32	<0.001
Multivariable	LAT, per 10 ms	0.99	0.94–1.04	0.690
Multivariable	DSC, per 0.1 increase	1.70	1.14–2.54	0.009

Key Points:

• Persistent atrial fibrillation (PsAF) remains a significant clinical challenge, often marked by high rates of post-ablation recurrence. While endo-epicardial dissociation is a known factor in the PsAF substrate, the underlying mechanisms within the 3D atrial wall that lead to this phenomenon remain incompletely understood.

• In this study, we utilized personalized digital twin technology derived from LGE-MRI scans of 21 patients to investigate the association between transmurally heterogeneous fibrosis and endo-epi electrical dissociation.

• Our findings reveal that 3D transmurally heterogeneous fibrosis architecture is associated with endo–epicardial electrical dissociation in PsAF. In multivariable analysis, fibrosis percentage, transmembrane voltage difference, and DSC were independently associated with reentrant activity.

• This work not only advances the mechanistic understanding of PsAF arrhythmogenesis but also suggests that successful clinical ablation strategies must account for the full 3D complexity of the atrial wall to target hidden intramural and epicardial pathways.

Data Availability Statement

All data generated or analyzed during this study are available from the corresponding author on reasonable request.