This retrospective study was approved by the institutional review board and included 16 patients with ARVC. Table 1 provides demographic information for this cohort. All patients were adults with a median age of 31 years, and 75% of them were female. All had VTs induced during clinical EPS, where information about the patients’ VTs were recorded. Genetic testing for ARVC risk variants was performed for the entire cohort. Eight patients were found to have PKP2 loss-of-function variants, while the remaining eight did not test positive for variants in any of the known causal genes and were thus considered gene-elusive (GE; James et al., 2021).

All 16 patients had RV enhancement on LGE-CMR; none of them exhibited enhancement on the left ventricle (LV). The clinical parameters previously found to be associated with VT in ARVC (RV ejection fraction and RV end-diastolic volume index [RVEDVI]) (te Riele et al., 2014) did not differ significantly between the GE and PKP2 groups. There was also no statistically significant difference in any of the other common clinical features between the two genotype groups.

Our Geno-DT approach incorporates, in the patient-specific computational models, cellular EP that is distinct between the two genotypes, PKP2 and GE. The cell-level electrical behavior for the GE genotype was modeled based on the ten Tusscher and Panfilov, 2006 (TT2) human ventricular model (see Methods). Since there is currently no cell model representation of PKP2 pathogenic variant that can be used in organ-level simulation in terms of computational tractability, we developed one here, by modifying the GE cell model to represent remodeling in the sodium currents and calcium-handling as reported in experimental studies (Sato et al., 2009; Kim et al., 2019; Lyon et al., 2021). Details on the PKP2 cell model implementation can be found in Materials and methods.

The new PKP2 model includes a downscaled maximum conductance for the sodium current (I_Na), resulting in a peak inward sodium current at a membrane potential of –36 mV that is truncated from –502.71 pA/pF in the GE model to –171.56 pA/pF in the PKP2 model (Figure 2A). Additionally, the PKP2 cell model includes various calcium currents, each modified from the baseline GE model (see Methods). These changes collectively produce an altered calcium transient (CaT) with a larger area under the curve and a faster decay, giving it a more acute shape (Figure 2B). The CaT in the PKP2 cell model has a longer time to peak (32 vs 27 ms), a significantly higher peak (0.0023 vs 0.00084 mM), and a shorter time to 90% return from peak (319 vs 379 ms).

The resulting action potential (AP) shape of the PKP2 cell is distinct from that of the GE cell, as depicted in Figure 2C. The downregulation of I_Na results in a lower upstroke peak and slower maximum upstroke velocity (V_max). The changes to calcium handling led to an elevated resting membrane potential and longer AP duration at 90% depolarization (APD₉₀). These differences in AP behavior between the two cells underlie the organ-level EP differences between PKP2 and GE genotypes.

Since in this study we used a rapid pacing protocol to induce VTs in the personalized ARVC models, cell-level restitution properties, and specifically the difference in restitution properties between the PKP2 and GE cell models, have important consequences for ARVC arrhythmogenesis. Figure 3 examines these restitution properties. The GE cell model exhibited a decrease in APD₉₀ with diastolic interval (DI) shortening (Figure 3A), consistent with normal calcium-handling behavior (Franz, 2003). In contrast, the PKP2 cell model’s AP showed some shortening of phase two duration but with very little reduction in plateau amplitude, resulting in more triangular APs at fast pacing without much change in total duration.

Figure 3B highlights the differences in restitution curves for the PKP2 and GE cell models. Notably, at the fast-pacing intervals of the steep initial phase (25–160 ms), the two restitution curves diverge significantly. While the GE cell APD₉₀ continued to decrease to 148ms before loss-of-capture, the PKP2 cell APD₉₀ only decreased to 208 ms. The maximum slopes of the restitution curves were measured to be 1.16 and 0.59 for the GE and PKP2 cell models, respectively. The substantially lower slope for the PKP2 cell’s restitution curve indicated a reduced ability of its AP to adapt to fast pacing rates.

The GE and PKP2 cell models described above were incorporated in LGE-based geometrical models of the patients’ ventricles from the respective patient groups. In addition to EP remodeling, ARVC is also characterized by structural remodeling, i.e the presence of dense scar and diffuse fibrosis. While replacement scar is non-conductive, myocardium in the diffuse fibrosis regions exhibits altered EP properties, which were represented in the personalized ventricular models by additional alterations in cell-level properties in these regions, as we have done previously (Shade et al., 2020; O’Hara et al., 2022); see Materials and methods for detail. The resulting heart models of ARVC patients thus represented both EP and structural remodeling in ARVC.

Here, we first utilized the LGE-reconstructed ventricular geometrical models to examine the distribution of structural remodeling in various regions of the RV and its relationship to VT. We found that across the entire ARVC cohort, the diffuse fibrosis was most abundant at the base of the RV and least abundant at the RV apex (Figure 4A). There were significant differences in the amount of diffuse fibrosis between the basal, mid, and apical regions (14.01 ± 4.37% for basal RV vs 5.41 ± 2.61% for mid RV vs 1.13 ± 1.71% for apical RV, p < 0.0001 for pairwise comparisons). A similar trend was observed for dense scar amounts with significant differences between the basal, mid, and apical regions (5.51 ± 2.71% vs 1.74 ± 1.99% vs 0.42 ± 0.90%, p < 0.05 for pairwise comparisons; Figure 4B). There was no significant difference when comparing the amounts of structural remodeling between anterior, lateral, and posterior regions.

We found that the distribution of diffuse fibrosis was different in the two patient groups, with GE patients having a significantly higher amount of diffuse fibrosis burden than PKP2 patients in the basal anterior wall (6.07 ± 3.18% vs 1.85 ± 0.017%, p < 0.01 for pairwise comparisons) and basal lateral wall (5.77 ± 2.34% vs 3.24 ± 1.40%, p < 0.05 for pairwise comparisons), and a significantly lower amount in the basal posterior wall (3.36 ± 2.92% vs 7.73 ± 4.38%, p < 0.05 for pairwise comparisons; Figure 4C). There was no statistical significance in the dense scar distribution between the two genotype groups.

Next, we performed simulations, with the personalized Geno-DT heart models, of VT induction following rapid pacing (see Materials and methods) and evaluated the induced VT circuits in the entire cohort. We performed a total of 144 ventricular simulations (16 patients ✕ 9 pacing locations) to determine the induced VT circuit morphologies. We examined the correlations between the structural remodeling features (diffuse fibrosis volume [DF], dense scar volume [DS], total fibrotic remodeling volume (both scar and diffuse fibrosis) [TFV]) and the number of unique VT morphologies induced during EPS across the different RV AHA segments. Figure 5 summarizes pairwise correlations between these variables for both GE and PKP2 patient groups. Our analysis revealed a highly positive correlation between the number of VTs induced during EPS and the volume of each type of structural remodeling (both DF and DS) and total fibrotic volume (TFV) in the GE patient group (r = 0.91, 0.91, 0.92 for DF, DS and TFV respectively), while in the PKP2 group, we observed only moderate correlation (r = 0.65, 0.7, 0.7 for DF, DS and TFV respectively).

Importantly, we found a strong correlation between VTs induced during EPS and those induced in simulations within the different RV regions in both patient groups (r = 0.99 for both groups), indicating that our heart digital twins effectively capture the occurrence pattern of EPS-induced VTs. Notably, although the structural remodeling in the PKP2 group was less correlated with EPS VTs as compared to the GE group, the VTs induced in the PKP2 models still exhibited high correlation with the VTs recorded during EPS.

These findings suggest that structural remodeling plays a more important role in VT arrhythmogenesis in GE patients than in PKP2 patients. The moderate correlation in the PKP2 group indicates that additional factors contribute to VT, specifically the EP remodeling associated with the PKP2 pathogenic variants.

We next assessed the predictive capability of the Geno-DT approach in determining accurately the VT circuit numbers and locations in the two ARVC genotype groups. In Figure 6, the bullseye plots summarize the number of unique VT morphologies induced during EPS and those induced in Geno-DTs for each RV AHA segment in all 16 ARVC patients. In the GE patient group, Geno-DT resulted in 28 distinct VTs being induced, 25 of which were observed during EPS. The PKP2 group showed comparable results, with 25 VT morphologies induced in PKP2 ARVC models and 25 reported in the EPS record. In terms of VT circuit locations, Geno-DT captured all VT locations that were observed during EPS except for one VT on the mid lateral RV wall in the GE group. In the PKP2 group, Geno-DT predicted all the VT locations recorded during EPS, except for one VT on the basal anterior RV wall and two on the basal posterior wall. Hence, in both cases, Geno-DT accurately predicted the locations of nearly all the VTs induced during EPS.

Additionally, for each individual ARVC patient, we compared the AHA segment of every VT circuit induced in each Geno-DT to the patient’s EPS VT record obtained from EPS. In Table 2, we summarized the capability of Geno-DT in predicting VT locations. The results showed outstanding accuracy, sensitivity, and specificity in both GE and PKP2 patient groups.

Figure 7A and B illustrate the above findings with two examples, one from GE and one from PKP2 group. For the GE patient (Figure 7A), the genotype-matched model induced three VT morphologies: two figure-of-eights on the anterolateral and posterolateral RV wall and a single-reentry VT on the posterior RV wall. This GE patient had two VT circuits observed during the clinical EPS, each one on the basal and mid lateral wall of the patient’s RV. Although the clinical EPS report did not mention VT induced on the posterior RV wall, it did mention that fractionated potentials, which are considered as VT circuit indicators, were observed in the posterior RV, and thus, clinical ablation was done there. Therefore, we consider all our predicted VTs in this patient in agreement with the clinical findings. Similarly, for the PKP2 patient, the genotype-matched model induced five VTs: a figure-of-eight reentry on the anterolateral wall, a single-reentry VT and two figure-of-eight VTs on the lateral RV wall, as well a single-reentry VT on the posterior RV wall (Figure 7B). This PKP2 patient underwent two clinical ablation procedures within one year. The first ablation identified three VTs, one in the mid lateral region, one in the basal posterolateral region and one in the basal posterior region. The EPS procedure during the second ablation identified two sustained VTs on the anterolateral and lateral wall. These examples showcase the excellent correspondence between Geno-DT VT circuit predictions and clinical observations.

Once we ascertained the ability of Geno-DT to correctly predict VT circuits in ARVC patients, we investigated the contribution of pathogenic variants (via the corresponding genotype-specific EP properties) in creating propensity to VT. To do so, we switched the cellular models between GE and PKP2 patient groups to introduce a mismatch between personalized structural remodeling and genotype-specific EP properties. We then repeated the simulations under these mismatched conditions. Figure 8C provides a summary of the findings. We also present examples in Figure 8A and B, which are the same two patients’ ventricular models included in Figure 7 but simulated under genotype-mismatched conditions. Evident from the figure is that VT reentrant circuits were also induced under genotype-mismatched conditions, however, they are distinctly different from those in the genotype-matched models, and very different from the VTs induced during EPS. In Figure 8A, the mismatched GE DT had only one VT circuit induced on the posterolateral RV wall, and in Figure 8B the mismatched PKP2 DT presented only one VT in the anterolateral RV wall. For both the GE and PKP2 examples, the genotype-mismatched DTs missed many VT circuits induced during EPS. The bullseye plots in Figure 8C emphasize the underprediction of the genotype-mismatched Geno-DTs, in comparison to those in Figure 6.

To gain a better understanding of the mechanisms underlying the induction of different reentrant circuits resulting from EP remodeling, we analyzed in detail the VT circuits simulated under genotype-matched and genotype-mismatched conditions. Two examples of this analysis are presented in Figure 9 for the PKP2 ventricular model from Figure 7B. Animation of the VT circuits propagation is available in Figure 9—video 1.

Reentrant circuit 1 is the VT induced in the lateral RV wall of the PKP2 (genotype-matched) DT (Figure 9A). This circuit involved two dense scar islands surrounded by diffuse fibrosis, creating an isthmus of low conductivity in between. The initial wavefront elicited from a basal posterior pacing site propagated around the dense scar and through the isthmus. As the exiting wavefront encountered non-fibrotic myocardium, it had a slow conduction due to the lowered upstroke velocity and elevated resting membrane potential resulting from PKP2 EP remodeling. Consequently, downstream tissue failed to be excited due to source-sink mismatch. Two other wavefronts propagated from either side of the dense scar islands into this unexcited myocardium and collided. In the time it took for the two opposing wavefronts to meet, the non-fibrotic myocardium at the entrance of the isthmus remained refractory (PKP2 cell model has extended refractoriness), thereby allowing the collided wavefront to travel back up the isthmus, meet refractory tissue and ultimately form the figure-of-eight reentry. In contrast, no reentry was induced at this same location with the incorporation of GE EP properties (Figure 9B). Specifically, as the wavefront left the isthmus and encountered non-fibrotic myocardium, its propagation speed was higher due to the GE cell model’s rapid upstroke velocity and unchanged resting membrane potential. Because of the fast conduction velocity, the isthmus did not have time to recover, thereby excitation occurred only of tissue ahead of the wavefront.

Reentrant circuit 2 was induced in the posterior RV wall of the PKP2 genotype-matched DT (Figure 9C). The wavefront elicited from the same basal posterior pacing site had slowed down significantly while traveling through a band of diffuse fibrosis. In areas where the band was wide, the conduction velocity decreased to the point where the wavefront was unable to excite non-fibrotic myocardium due to source-sink mismatch. This was a result of the fact that the PKP2 cells have slowed upstroke velocity and elevated resting membrane potential. An adjacent wavefront was thereby able to turn into this unexcited region and initiate reentry. In contrast, no reentry occurred at this location in the GE DT (Figure 9D). The GE cell’s rapid upstroke velocity allowed the wavefront to excite the non-fibrotic myocardium after exiting the wide diffuse fibrosis band. Consequently, the wavefront leaving the diffuse fibrosis was more planar, avoiding source-sink mismatch, and did not result in reentry.

Together, these results help illustrate the contribution of genotype-based EP differences to the variability of organ-level wavefront propagation patterns.

This retrospective study cohort included 16 patients from the Johns Hopkins ARVC database who were diagnosed with ARVC based on 2010 Task Force Criteria between 2010 and 2018. Patients were included if they (1) underwent genetic testing for ARVC risk variants that demonstrated either no pathogenic variants (GE patients) or pathogenic PKP2 variant (PKP2 patients), (2) had inducible VT during EPS, and (3) had LGE-CMR demonstrating the presence of RV fibrosis. Clinical reports from EPS were reviewed to identify both the number of distinct VT morphologies and the locations of induced VT within the RV. Patients’ CMR images were reviewed clinically to identify the presence/absence of RV enhancement (S.Z.). Patient clinical characteristics are shown in Table 1.

The different EP properties of cardiomyocytes in the GE and PKP2 genotype groups were represented with two different cell ionic models. For non-fibrotic myocardium regions of the GE group, we used the ten Tusscher and Panfilov, 2006 model with the addition of a late sodium current representation (O’Hara et al., 2011). A late sodium current (I_NaL) formulation was added, as done in our previous studies (Cartoski et al., 2019; Shade et al., 2020; Arevalo et al., 2016; Prakosa et al., 2018; Shade et al., 2021). For non-fibrotic myocardium regions of the PKP2 group, we developed a new PKP2 model by modifying the GE model based on experimental data (Lyon et al., 2021 and Sato et al., 2009). These modifications are summarized in Figure 10. Our resulting PKP2 model recapitulates the altered sodium current dynamics and calcium-handling properties observed in experimental recordings of cardiomyocytes from ARVC hearts.

In our PKP2 ionic model, the maximum channel conductance for sodium current (I_Na) was decreased by 70% to reflect reported experimental current dynamics from Sato et al., 2009. The calcium-handling in our PKP2 model was adjusted to reproduce experimental data from Lyon et al., 2021, which reported a decreased L-type calcium current (I_CaL) and higher amplitude calcium transient (CaT). We scaled the maximal conductance of I_CaL by 50%, reducing its peak current (9.03 vs 6.84 pA/pF) proportional to the experimental values. To counteract the accompanying decrease in CaT amplitude produced by this modification, we also upregulated the background calcium current (I_bCa), representative of connexin 43 (Cx43) hemichannel-mediated calcium entry, by fivefold. This upregulation was in line with the in vitro properties of PKP2-deficient cells: disrupted cell-cell adhesion, functional dysregulation of Cx43, increased membrane calcium permeability, and excess free calcium concentrations.

Studies from Lyon et al. have further shown that PKP2-deficient cardiomyocytes have reduced ryanodine receptor (RYR2) channel expression but enhanced channel sensitivity to calcium. We incorporated reduced channel expression as decreases in maximal rate of calcium movement for both the RyR2-mediated SR release (0.102 vs 0.0816 ms-1) and leak (3.61E-4 vs 2.88E-4 ms-1) currents. Since the baseline TT2 model did not represent individual RyR2 channels, we indirectly accounted for increased sensitivity by increasing free calcium concentrations in the SR (CaSR) and subspace (CaSS). CaSR overload was achieved by decreasing calsequestrin affinity as a 40% reduction in the half-saturation constant for sarcoplasmic buffering (0.3 vs 0.18 mM). CaSS was increased from the reduced calcium diffusion gradient between the subspace and the cytosol; a 20% downregulation of the sodium-calcium exchanger current (I_NaCa) further contributed to this reduced gradient. The enhanced calcium concentration of all three compartments corresponded with the experimentally measured calcium levels for PKP2-deficient cardiomyocytes and effectively shifted the concentration vs RyR2-binding curve leftward.

To create the action potential duration (APD) restitution curves of the GE and PKP2 models, we used a standard S1-S2 stimulus protocol similar to that described in ten Tusscher and Panfilov, 2006. Twenty S1 stimuli were applied at a basic cycle length (BCL) of 1 Hz to allow the models to reach a steady state. A single S2 extra stimulus of twice the diastolic threshold was then delivered at some diastolic interval (DI) following the action potential generated by the last S1 stimulus. The model was allowed to retain steady state (5 beats at BCL) before another S2 stimulus was applied at a decreased DI. This step was repeated for DIs decreasing from 1000 ms to 50ms in intervals of 25ms. An APD restitution curve was generated by plotting the APDs at 90% repolarization generated by the S2 stimuli against preceding DIs.

For regions of diffuse fibrosis, we used the TT2 model with modifications based on data reported by Coppini et al., 2013. The EP properties of this modified model have been validated in previous studies (Shade et al., 2020; O’Hara et al., 2022). Maximal conductance of I_NaL and I_CaL were respectively increased by 107% and 19%; maximal conductance of I_Kr, I_Ks, I_to, and I_K1 were respectively decreased by 34%, 27%, 85%, and 15%; sodium-calcium exchanger activity was upregulated by 34%; sarcoplasmic/endoplasmic reticulum calcium ATPase activity was downregulated by 43%.

Geometrical model of each patient’s ventricles was reconstructed based on the patient’s 2D LGE-CMR scans. Each diastolic short-axis image stack had a median of 11 slices with an average of 9.58±2.16 mm slice thickness and average in-plane axial resolution of 1.56±0.49 mm. Both the RV and LV myocardium were segmented from the LGE-CMR using a semi-automatic function of CardioViz3D which has been previously validated by our team (Arevalo et al., 2016; Shade et al., 2020; Cartoski et al., 2019). This segmentation method utilized landmarks to define the endocardial and epicardial boundaries.

Unlike our previous digital twin studies in ischemic cardiomyopathy (Arevalo et al., 2016; Shade et al., 2020; Cartoski et al., 2019), patients in the current cohort were right-dominant ARVC patients with only RV fibrosis detectable on LGE-MRI. We used Otsu thresholding to binarize the segmented myocardium into high- and low-intensity regions. The mean value and the standard deviation (SD) of the low-intensity region was calculated. We used the mean value of the low-intensity region as the reference mean of non-fibrotic myocardium and an intensity of ≥4 SD above the reference mean was classified as the dense scar region. Voxels between 2 and 4 SDs above the reference mean were classified as diffuse fibrosis; all other voxels were labeled as non-fibrotic myocardium (Zhang et al., 2021).

After identifying all different tissue types, we generated high-resolution finite-element tetrahedral meshes with an average resolution of 398 μm from the segmentation using finite-element analysis software (Mimics Innovation Suite; Materialise, Leuven, Belgium), as done previously (O’Hara et al., 2022). We then used a validated rule-based method to incorporate fiber orientations to each element of the computational mesh (Bayer et al., 2012).

We used the openCARP software package to simulate electrical activity in monodomain representation of the myocardium (Plank et al., 2021). Full details regarding the simulation of electrical activity in the heart digital twins can be found in previous publications (O’Hara et al., 2022; Prakosa et al., 2018; Sung et al., 2020; Shade et al., 2020). Each heart digital twin was paced sequentially from 9 uniformly distributed endocardial RV locations (see Figure 1) using a validated rapid pacing protocol described in our previous studies (Prakosa et al., 2018; Arevalo et al., 2016; Cartoski et al., 2019). At each pacing site, six pacing stimuli (S1) were delivered at a 600ms cycle length, followed by a premature stimulus (S2) delivered 300ms after S1. If S2 did not result in reentrant arrhythmia, we shortened the S1-S2 interval in 10ms steps until arrhythmia was induced or the S2 failed to capture the tissue. If arrhythmia was not induced, we delivered additional stimuli (S3 and S4) in the same way as S2. Induced VT in the Geno-DT models was defined as re-entry after sustaining at least 2 cycles at the same critical site, as in our previous studies (Sung et al., 2020; Shade et al., 2021).

After inducing the re-entries, we analyzed activation maps to identify the VT morphologies and locations. Reentries induced from different pacing sites but occurring in the same location with the same morphology were classified as repetitive. Only unique VT circuits were counted. Simulations were conducted blind to the clinical data.

Our study has a small sample size, limited by the fact that ARVC is relatively a rare disease. Moreover, since the goal of this study was to develop an image-based heart digital twin approach to accurately predict VT circuit locations, LGE-CMR scans and EPS records were needed to reconstruct the model and validate our predictions respectively; only a few patients in the database had both. Additionally, our study focused on patients with only RV enhancement since research has shown that most clinical VT ablation sites are on the RV for ARVC, even for those with LV involvement (Marchlinski et al., 2021). Furthermore, we considered only two genotypes, GE and PKP2, as human experimental data on cellular EP properties for other pathogenic variants implicated in ARVC (e.g., DSP, DSG-2), from which to construct cell models, is lacking; in addition, because the prevalence of these causal variants is lower compared to that of PKP2, there were not enough patients in the database with the corresponding clinical data needed for model construction and validation (clinical images, EPS study, etc.).