Identifying the Presence and Characteristics of Mid-myocardial and Epicardial Fibrosis from Intracardiac Electrograms in Patients Undergoing Ventricular Arrhythmia Ablation Using a Transformer-based Self-supervised Classifier

Abstract

Background:

Catheter ablation is an essential tool for ventricular arrhythmia (VA) management, yet sustained procedural success is hindered by the limited ability to identify nonendocardial arrhythmogenic substrates during the procedure. Although delayed enhancement cardiac MRI is the reference standard for detecting myocardial fibrosis, barriers including cost, workflow complexity, and artifacts in patients with implantable devices limit its pre-procedural use. We hypothesized that intracardiac electrograms provide sufficient information to infer scar beyond the endocardial surface and that this information can be harnessed by machine learning techniques.

Methods:

This retrospective study included a total of 131,584 electrogram (EGM) signals collected from 46 patients undergoing ventricular arrhythmia ablation. A stratified patient-wise split was employed to create the training/validation set (N = 37) and the testing set (N = 9), while ensuring similar distribution of scar types. We developed a novel imaging processing workflow to create scar labels using coregistered cardiac MRI and electroanatomic mapping surface meshes. We developed a transformer-based self-supervised model EGM2Scar-AI alongside basic convolutional neural network models using either EGM waveforms or EGM-derived short-time Fourier transform (STFT) spectrograms.

Results:

Average task-specific AUROC of both the basic and STFT-based CNN are 0.729 [0.725–0.732] and 0.729 [0.726–0.733] respectively, while EGM2Scar-AI performed significantly better with an average AUROC of 0.822 [0.819–0.825] across all three scar types. All models perform better on endocardial and mid-myocardial fibrosis identification with modest reduction in performance for epicardial fibrosis. Sensitivity improved significantly with a transformer-based architecture without appreciable changes in specificity.

Conclusions:

Our study demonstrates that routine intracardiac electrograms enable the identification of endocardial, mid-myocardial, and epicardial scar using a transformer-based self-supervised deep learning model. Ultimately, our model has the potential to provide MRI-like fibrosis maps during EP procedures without the need for external imaging and increase dataset size for patient-specific models in ventricular arrhythmia research.

Graphical Abstract

Introduction

Ventricular arrhythmia (VA) ablation has become a cornerstone therapy for reducing implantable cardioverter-defibrillator (ICD) shocks and preventing electrical storm.¹ Despite procedural advances, long-term success remains limited by incomplete identification and targeting of arrhythmogenic substrate.² Current mapping techniques primarily characterize the endocardial surface, leaving the distinction of midmyocardial or epicardial fibrosis difficult in most patients.³ Late gadolinium-enhanced cardiac magnetic resonance imaging (LGE-CMR) has been validated against histopathology as the gold standard for fibrosis detection.⁴ However, routine integration of LGE-CMR into ablation workflow is limited by cost, accessibility, artifact due to implanted devices or poor breath holding, contrast allergy or renal dysfunction, and difficulty of coregistration between images and electroanatomic maps during the procedure. Consequently, the ability to visualize deep myocardial fibrosis at the point of ablation remains an unmet clinical need.

Advances in ablation technology, such as deep-penetrating radiofrequency, ultra-low temperature cryoablation, pulsed field ablation, noninvasive radioablation, and VINTAGE, provide an opportunity to treat deep substrate.^5–7 Real-time access to structural information during the procedure could enhance substrate-directed ablation, minimize unnecessary lesions, and improve procedural outcomes. Cardiac contact electrograms (EGMs) capture detailed local depolarization and repolarization characteristics that reflect underlying myocardial structure. Our prior work has demonstrated that intracardiac monophasic action potential recordings contain physiologic information predictive of sudden cardiac arrest risk.⁸ We have also shown that machine learning applied to surface electrocardiograms can infer underlying mechanical and structural abnormalities.⁹

We hypothesized that a self-supervised model trained jointly on intracardiac bipolar and unipolar EGMs and surface ECGs could identify the local presence of mid-myocardial and epicardial fibrosis. Such a model could ultimately provide three-dimensional fibrosis maps using only signals obtained during routine intracardiac mapping, without the need for external imaging. It may also inform the development of patient-specific digital-twin models in patients with prior electrophysiologic mapping. In this study, we introduce EGM2Scar-AI, a transformer-based self-supervised model that predicts myocardial fibrosis distribution across the endocardial, mid-myocardial, and epicardial regions from contact endocardial electrograms. We analyze its performance in creating three-dimensional fibrosis maps including deep substrate against gold-standard LGE-CMR and investigate the characteristics of the signals related to distributions of fibrosis.

Methods

Study population and setting

We conducted a retrospective cohort analysis of 190 patients who underwent both ventricular arrhythmia ablation and CMR imaging within 3 years of the ablation procedure between January 1, 2013, and January 1, 2023, at Stanford Health Care. Patients were included if a compatible electroanatomical map (EAM) was available with an LV endocardial map collected using CARTO 3^™ electroanatomic mapping system (Biosense Webster, Diamond Bar, Ca, USA), CMR quality was sufficient for scar localization, and adequate anatomic structures were available on the EAM for coregistration. Both mapping and ablation catheters were used to collect the signals, including the PENTARAY^™, DECANAV^™, and ThermoCool SmartTouch^™ catheter (Biosense Webster, Diamond Bar, Ca, USA). EGM signals were collected during sinus or paced rhythm, not during ventricular tachycardia. A total of 46 patients undergoing ventricular arrhythmia ablation procedure were included for model training and testing (See Consort diagram, Figure 1). Fibrosis-labeled EGM signals from all patients (N=46) were split at the patient level into a training/validation set (37 patients, 80%) and a held-out test set (9 patients, 20%). We performed a stratified group split, ensuring similar distributions of binary fibrosis location labels (endocardial, mid-myocardial, epicardial) between the training/validation and test sets. The split was determined a priori using only patient-level fibrosis labels without electrogram features, to avoid information leakage. All electrograms from a given patient were assigned to either training/validation or testing set. A total of 131,584 electrograms were used for model development and testing (Table 1). The Stanford University Institutional Review Board approved this retrospective study (IRB-60644). The requirement for informed consent was waived due to retrospective use of de-identified data. The data that support the findings of this study are available from the corresponding author upon reasonable request.

Figure 1: Study Cohort Selection and Data Partitioning for EGM2Scar-AI Model Development.

Consort showing patient selection for model development and testing. Among 190 patients undergoing ventricular arrhythmia ablation with cardiac magnetic resonance imaging performed within 3 years, 106 were excluded because a compatible electroanatomic mapping export was not available. Of the remaining 84 patients, 38 were excluded because of absent left ventricular mapping, unavailable delayed gadolinium enhancement imaging, or corrupt electroanatomic mapping data. The final cohort included 46 patients used for EGM2Scar-AI model development and evaluation. Patients were split at the patient level into a train/validation cohort of 37 patients comprising 101,591 electrograms and an independent testing cohort of 9 patients comprising 29,993 electrograms. *cMRI, cardiac magnetic resonance imaging; DGE, delayed gadolinium enhancement; EAM, electroanatomic mapping; VA, ventricular arrhythmia; LV, left ventricular.

Table 1:

Patient characteristics

	Train/Validation (N = 37)	Test (N = 9)	p-value
Age (years, ±sd)	54.9 ± 17.0	52.6 ± 1.4	0.70
Number of points (IQR)	1029 (236–2564)	1774 (367–3256)	0.86
Endocardial scar points (±sd)	608 (2780)	634 (985)	0.44
Mid-myocardial scar points (±sd)	811 (3469)	586 (821)	0.13
Epicardial scar points (±sd)	856 (3544)	837 (1052)	0.11
Number of MRI short axis slices (±sd)	10.8 ± 2.2	10.4 ± 2.2	0.62
Male (%)	20 (54)	6 (67)	0.49
VT (%)	23 (62)	7 (78)	0.38
PVC (%)	14 (38)	2 (22)	0.38
CAD (%)	8 (22)	3 (33)	0.46
MI (%)	5 (14)	3 (33)	0.16
AF (%)	10 (27)	2 (22)	0.77
CMR scar (%)	15 (41)	3 (33)	0.69
EF (percent, ±sd)	49.7 ± 12.3	46.8 ± 14.0	0.54
ICD (%)	10 (27)	2 (22)	0.77
Antiarrhythmic drug (%)	23 (62)	7 (78)	0.38

VT, ventricular tachycardia; PVC, premature ventricular contraction; CAD, coronary artery disease; MI, myocardial infarction; AF, atrial fibrillation; CMR, cardiac magnetic resonance; EF, ejection fraction; ICD, implanted cardioverter-defibrillator. P-values were derived from chi-squared test for categorical variables, Student’s t-test for continuous variables with a normal distribution, and Mann-Whitney U-test for continuous variables with a non-normal distribution.

CMR data acquisition and fibrosis model creation

Cardiac MRI images were acquired using the GE Medical Systems platform (GE Medical Systems, Waukesha, Wisconsin) with magnetic field strengths of either 1.5T or 3.0T. LGE sequences were acquired following standard institutional gadolinium-based contrast administration and delayed imaging protocol. MRI images were collected using GE Signa HDxt 1.5T, Signa Explorer 1.5T, and Signa Premier 3.0T. Imaging data was exported from the institution’s research information technology imaging data warehouse and stored in a protected health information-secure server.

Patient-specific three-dimensional cardiac meshes were created using short axis LGE slices to identify the three-dimensional extent of fibrosis as determined by delayed gadolinium enhancement. The short axis in plane resolution ranged from 256×256 to 512×512 voxels depending on the LGE-CMR protocol. Anatomic location of myocardial fibrosis is identified using the clinical reports created by board certified cardiovascular and radiologic imaging experts. Fibrosis voxel annotation was done with blinding to electrogram or clinical case information. We employed a semi-automated approach combining a deep learning algorithm with visual inspection to generate 3D surface meshes from raw MRI images¹⁰ (Supplemental Methods and Figure S1). The result was a three-dimensional model with a myocardial volume spanning the endocardial and epicardial surface. Within the myocardial volume, regions are denoted as endocardial, mid-myocardial, epicardial fibrosis, or no fibrosis. (Figure 2, right).

Figure 2: Multimodal Data Integration Workflow for Electrogram-Based Prediction of Transmural Myocardial Scar.

Electroanatomic mapping data, including unipolar, bipolar, and reference electrograms, were exported from the clinical mapping system (left). Late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) imaging was processed to generate a 3D myocardial surface mesh and corresponding fibrosis across endocardial, mid-myocardial, and epicardial layers (top right). The electroanatomic map and CMR-derived geometry were then coregistered to spatially align intracardiac sampling points with tissue labels (center). Each electrogram was paired with its corresponding myocardial layer and scar classification to form supervised model inputs. Deep learning models used these paired inputs to predict the presence of endocardial, mid-myocardial, or epicardial scar at each mapped location (bottom right).

Electroanatomic map and ventricular contact electrogram

EGM waveforms, recording points, and anatomical mesh information were exported from the CARTO 3^™ electroanatomic mapping system (Biosense Webster, Diamond Bar, CA, USA) as zipped archives containing .xml, .mesh, and .txt files. Anatomical meshes of the endocardial mapping information from each ablation study were processed and exported from the archives and stored as object files. We developed custom software in Python to extract waveforms and cartesian coordinates associated with each collected left ventricular point. Each EGM point contains reference (surface ECG), local unipolar, and local bipolar signals (Figure 2, left). A total of 2500 voltage samples were collected per channel for each EGM point using a sampling frequency of 1000 Hz. After extraction, point and waveform data was stored for model training as NumPy arrays.

A zero-phase, 4th order Butterworth bandpass filter was applied to all signal channels to decrease high frequency noise (0–100 Hz). Each channel of individual EGMs is scaled to an absolute peak value of 1 due to high variability between channels and loss of convergence when absolute voltages are used. All three channels are collated into a 3×2500 NumPy array. A Short-time Fourier transform (STFT) is applied to generate spectrograms using 3 different combinations of frame lengths and hop lengths specifically for the development of the STFT model.¹¹ The basic and transformer models used preprocessed EGM waveforms rather than spectrograms as input.

Coregistration

Coregistration of the EAM mesh and whole heart MRI chamber segmentation were performed manually using Blender 4.2 (Blender Foundation, Amsterdam, NL). Custom Python software was created for use in Blender to export the transfer matrix that coregisters the EAM mesh to the MRI segmentation. See Supplemental Methods for Python code. During coregistration, whole heart MRI chamber segmentations did not include information about scar to blind coregistration to scar labels. Similarly, EAM meshes did not include point locations, waveform characteristics, or activation times. Coregistration was only attempted if at least one other cardiac structure was available in addition to the left ventricular cavity (such as the RVOT, aortic root, coronary sinus, or right atrium) to align the ventricle around its apex-to-base axis similar to prior studies.¹² Figure S2 shows an example of the coregistration process in Blender. The output from the coregistration step is a patient specific model that pairs contact electrograms with local fibrotic substrate for model development.

Fibrosis label creation

First, each EGM point is projected to the closest vertex on the endocardial surface of the processed LGE-CMR mesh. An orthogonal projection is performed from endocardial surface to epicardial surface to identify the myocardial region of interest and its endocardial, mid-myocardial, and epicardial sections (⅓ of the thickness for each section)^13,14. To reduce errors due to complex geometry of the myocardium, we removed EGM points near the basal LV region and left ventricular outflow tract. This is because these structures are poorly captured on CMR and don’t fit our assumptions of a smooth myocardial contour. A detection radius of 5 millimeters around the EGM point is chosen as a reasonable approximate area of influence detectable by the mapping catheter and delimited by the distance between CMR slices.^15,16 The scar mesh is then converted into a dense point cloud to facilitate localization. Labels for each EGM point are defined by the orthogonal projection at the nearest point on the endocardial surface. Endocardial, mid-myocardial, and epicardial scar is defined based on the distance of scar point from the endocardial surface to the epicardial surface divided into equal thirds. EGM points are plotted with both MRI and Carto meshes and provided in Figure 2 (center).

Development of Fibrosis Classifier Models

Generative AI assisted with code writing (Gemini 2.5 pro) using Python (v3.12)¹⁷. The code and results were reviewed by multiple human code reviewers. The transformer model was built and trained with Pytorch (v2.8.0+cu126) while CNN models were built and trained with TensorFlow (v2.19.0). Best models for each architectural design are used after hyperparameter tuning. Prediction of scar is treated as independent binary tasks at the final layer in each model.

Basic CNN

The baseline model consisted of a conventional 2D-CNN using residual network layers in TensorFlow. Convolutional layers were followed by fully connected (dense) layers and a final sigmoid activation function to generate probability estimates for each scar class.

Fourier transform-based CNN

A frequency-spectrum model was created using STFT applied to the EGM signals as described above. These spectrograms served as the model inputs. The architecture comprised a sequential two-dimensional CNN using standard convolutional layers. Prior work demonstrated that STFT features can improve ECG classification model performance over a 1D-CNN classier.¹¹

Transformer (EGM2Scar-AI)

The transformer model (EGM2Scar-AI) used an encoder-decoder framework with a patch embedding of 50 ms, followed by positional encoding, and stacked multi-head self-attention encoding layers. The model was pretrained on three masked reconstruction tasks (Figure S3).¹⁸ Pretraining employed a composite loss function consisting of voltage weighted mean absolute error (MAE) across different self-supervised tasks and a contrastive loss (NT-Xent) component.¹⁹ In the fine-tuning step for scar classification, the loss function is defined as the weighted binary cross entropy. The full transformer model architecture is shown in Figure 3A.

Figure 3: Transformer architecture, attention map and super-weight activation.

A) Transformer architecture of pre-training and classification tasks. B) Attention map of exemplar PC1 and PC2 waveforms based on averaged attention layer activation. C) Histogram of super-neuron scores, calculated by average activation across the entire data set multiplied by the output weight vector by each neuron. D) Exemplar electrogram patches associated with super neuron activations.

Statistics

All statistical analysis was conducted in Python (v3.12) using SciPy and Scikit-learn.²⁰ For patient characteristics, chi-squared test was used for categorical variables, Student’s t-test was used for continuous variables with a normal distribution, and Mann-Whitney U-test was used for continuous variables with a non-normal distribution. The primary measure of discrimination was the Area Under the Receiver Operating Characteristic curve (AUROC). Additionally, we calculated sensitivity, specificity, F1 score, positive predictive value (PPV), negative predictive value (NPV), and Area Under the Precision-Recall Curve (AUPRC) using standard formulations in alignment with the EHRA AI checklist.²¹ Decision thresholds for each model and each classification task (i.e. endocardial vs mid-myocardial vs epicardial) were determined independently using Youden’s J statistics²². 95% confidence intervals for all performance metrics were estimated using point-level bootstrapping with 1000 iterations and a sample size of 1000. Additional uncertainty estimates using patient-level resampling are described in Supplemental Methods. Principal component analysis and linear discriminant analysis were conducted for performance visualization of EGM2Scar-AI. The last feature space of the transformer classification head was extracted for each point in the training data set. The top two principal components of the 128-dimensional feature space were calculated. All points of the training dataset are then projected onto the top 2 principal components to generate a scatter plot for visualization. Class separation plots were generated corresponding to each scar classification task using linear discriminant analysis in scikit-learn. Spectral analysis of pre-training reconstruction tasks is conducted using Welch’s averaged periodogram method.

Results

Patient characteristics

Age, sex distribution, and arrhythmia ablation type (VT/PVC) were similar across groups, supporting generalizability of the model to different patient profiles. There were no differences in prevalence of comorbidities, including CAD, prior MI, AF, and CMR-defined scar or in rates of ICD implantation or arrhythmic drug use. Cardiac function was also similar, minimizing the risk that differences in ventricular remodeling confounded model performance.

A total of 131,584 EGM points were included in the study. Among those, the training and validation dataset included 101,591 points (77.2%), while the testing set included 29,993 points (22.8%) (Figure 1). The distribution of total number of EGM points and number of fibrosis points per patient are not significantly different between the training/validation and testing cohort. The presence of endocardial, mid-myocardial, and epicardial fibrosis are similarly distributed between the training/validation and testing cohort. EGM and CMR data richness was comparable. The total number of mapping points was similar between train/validation and test sets, indicating comparable mapping quality.(Table 1)

Classification performance of the basic CNN, STFT-based, and transformer model

AUROC of the basic and STFT-based CNN are similar, while the transformer model performed significantly better with an average AUROC of 0.822 (95% CI: 0.819–0.825). (Table 2, Figure 4B) All models perform better on endocardial and mid-myocardial fibrosis identification with modest reduction in performance for epicardial fibrosis. Sensitivity was similar with STFT optimization compared to the basic CNN model but improved significantly with a transformer-based architecture without appreciable changes in specificity. (Table 2) Receiver operating characteristic curves for endocardial, mid-myocardial, and epicardial classification tasks are shown in Figure 4A. Performance uncertainty estimates of the EGM2Scar-AI model using patient-level resampling are shown in Supplemental Table 1. Precision-recall curves for each classification task for each model are shown in Figure S6. Visualization of model performance using PCA plots shows appropriate clustering of true positive and true negative labels. There is significant overlap between true negative (red) and false negative (cyan) label clusters. PC1 explains 57.3% of variance and PC2 explains 17.6% of variance. (Figure 4C) Linear discriminant analysis demonstrates clearer separation between true and false labels for endocardial and mid-myocardial prediction tasks, but more overlap for epicardial classification. (Figure 4C)

Table 2:

Model Performance

	Sensitivity	Specificity	F1 Score	PPV	NPV	AUROC	AUPRC
Basic CNN
Endo	0.832 (0.823 – 0.842)	0.632 (0.625 – 0.638)	0.490 (0.481 – 0.498)	0.347 (0.339 – 0.354)	0.941 (0.938 – 0.945)	0.779 (0.774 – 0.785)	0.364 (0.353 – 0.374)
Mid	0.841 (0.830 – 0.850)	0.548 (0.542 – 0.554)	0.424 (0.416 – 0.433)	0.284 (0.277 – 0.291)	0.942 (0.938 – 0.945)	0.747 (0.740 – 0.753)	0.322 (0.311 – 0.332)
Epi	0.654 (0.643 – 0.664)	0.614 (0.608 – 0.620)	0.466 (0.458 – 0.475)	0.362 (0.354 – 0.370)	0.841 (0.836 – 0.847)	0.675 (0.669 – 0.681)	0.403 (0.392 – 0.414)
Average	0.787 (0.781 – 0.793)	0.574 (0.571 – 0.578)	0.459 (0.454 – 0.464)	0.324 (0.320 – 0.328)	0.912 (0.910 – 0.915)	0.729 (0.725 – 0.732)	0.358 (0.353 – 0.364)
STFT CNN
Endo	0.763 (0.752 – 0.773)	0.645 (0.639 – 0.651)	0.465 (0.457 – 0.474)	0.335 (0.327 – 0.343)	0.920 (0.917 – 0.924)	0.754 (0.747 – 0.760)	0.364 (0.353 – 0.375)
Mid	0.775 (0.764 – 0.785)	0.638 (0.632 – 0.644)	0.446 (0.437 – 0.454)	0.313 (0.306 – 0.320)	0.930 (0.926 – 0.934)	0.757 (0.750 – 0.763)	0.339 (0.329 – 0.350)
Epi	0.610 (0.599 – 0.621)	0.706 (0.700 – 0.712)	0.491 (0.482 – 0.500)	0.411 (0.402 – 0.420)	0.844 (0.839 – 0.849)	0.703 (0.697 – 0.710)	0.445 (0.434 – 0.457)
Average	0.754 (0.747 – 0.760)	0.583 (0.580 – 0.587)	0.448 (0.443 – 0.453)	0.319 (0.315 – 0.323)	0.901 (0.899 – 0.904)	0.729 (0.726 – 0.733)	0.384 (0.377 – 0.391)
Transformer
Endo	0.930 (0.923 – 0.936)	0.701 (0.696 – 0.707)	0.581 (0.572 – 0.589)	0.422 (0.414 – 0.431)	0.977 (0.975 – 0.979)	0.875 (0.872 – 0.879)	0.539 (0.526 – 0.552)
Mid	0.906 (0.898 – 0.914)	0.695 (0.689 – 0.701)	0.543 (0.534 – 0.552)	0.388 (0.379 – 0.397)	0.972 (0.970 – 0.975)	0.864 (0.859 – 0.868)	0.505 (0.491 – 0.520)
Epi	0.849 (0.840 – 0.857)	0.611 (0.605 – 0.617)	0.564 (0.556 – 0.572)	0.422 (0.415 – 0.430)	0.923 (0.919 – 0.927)	0.784 (0.779 – 0.789)	0.491 (0.479 – 0.503)
Average	0.834 (0.828 – 0.839)	0.689 (0.685 – 0.692)	0.549 (0.544 – 0.554)	0.409 (0.404 – 0.414)	0.941 (0.939 – 0.943)	0.822 (0.819 – 0.825)	0.498 (0.491 – 0.506)

Sensitivity, true-positive rate; Specificity, true-negative rate; F1 Score, harmonic mean of precision and recall; PPV, positive predictive value; NPV, negative predictive value; AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; Endo, endocardial scar; Mid, mid-myocardial scar; Epi, epicardial scar; CI, confidence interval; CNN, convolutional neural network; STFT, short-time Fourier transform.

Figure 4: Model performance.

A) ROC curve for endo-, mid-, and epi-cardial detection of EGM2Scar-AI. B) Overlay of ROC curve for any scar detection for CNN, STFT, and transformer models. C) PCA cluster plot (top), LDA plot (bottom) for endo-, mid-, and epi-cardial prediction tasks of EGM2Scar-AI.

EGM2Scar-AI model pretraining task performance

With respect to pretraining tasks, the transformer model successfully reconstructed masked portions of the ground truth EGMs with a high degree of fidelity. Example reconstructions from a single multi-channel EGM point are provided in Figure S4. Frequency-domain analyses comparing the original and reconstructed waveforms for a single EGM point (Figure S5A) and a population of 1,024 EGM points (Figure S5B) showed stable spectral preservation across reconstruction types, particularly within clinically relevant low-frequency bands. Additionally, average weighted MAE is similarly low across all tasks within the sample population consistent with pretrained model performance (Figure S5B).

EGM2Scar-AI model attention map and super-weight activation

The attention maps of the exemplar waveforms for the top 2 principal components of the output feature space are shown in Figure 3B. The model appropriately assigns higher attention scores to information-rich patches that are centered around QRS complexes (Figure 3B). The majority of neurons have low super-weight scores, while a small minority of neurons have scores above 20, which have an outsized impact on the performance of the model (Figure 3C). The exemplar patches that lead to the highest activation of the top 3 neurons and the bottom 3 neurons ranked by superneuron score are shown in Figure 3D. The top 3 neurons are activated by patches with synchronized deflections in the unipolar and reference channels. There is less change in scaled voltage over time of the bipolar waveform in these patches. The bottom 3 neurons are activated by patches without discernable synchrony in voltage variability over time between the 3 channels.

Case study of EGM2Scar-AI performance

A man in his 40s with ischemic cardiomyopathy, prior MI, and dual-chamber ICD implant presented for management of recurrent monomorphic VT despite antiarrhythmic therapy. LV endocardial mapping was performed to collect a total of 1,774 EGM points and ablation location included RV apical septum, LV apical septum, LV lateral apex and epicardial LV apex. Figures 5A and B show the RAO and LAO projection of coregistered electroanatomic map and LGE-CMR-derived myocardial surface meshes with scar indicated as yellow. Labels for endocardial, midmyocardial surface based on LGE-MRI segmentation are shown in Figure 5C (left) along with fibrosis distribution as predicted by EGM2Scar-AI (center) based on endocardial contact EGMs alone. For comparison, bipolar voltage-based scar defined by peak-to-peak voltage < 0.5 mV is shown in Figure 5C (right).

Figure 5: Patient specific fibrosis model generated from contact electrograms and relationship to bipolar voltage labels.

Coregistration of one patient’s LGE-CMR mesh in the right anterior oblique (A) and left anterior oblique (B) views showing endocardial and epicardial meshes with fibrosis displayed simultaneously with the electroanatomic map. C) Overlay of true labels fields (left), model predicted scar fields (center), and endocardial scar fields based on bipolar voltage criteria (right)

Discussion

In this study, we developed and validated EGM2Scar-AI, a transformer-based self-supervised model that predicts the presence and transmural distribution of myocardial fibrosis using endocardial contact intracardiac EGMs collected during routine electrophysiology mapping during ventricular arrhythmia ablation. By integrating information from local bipolar, unipolar, and surface ECG channels, the model accurately identified endocardial, mid-myocardial, and epicardial scar when benchmarked against LGE-CMR, with performance that exceeded both waveform- and STFT-based CNN models. These findings support the hypothesis that spatial and temporal relationships within contact EGMs encode information about deep myocardial structure that extends beyond conventional voltage criteria.

Our results build on prior work demonstrating that frequency-domain or advanced signal features can improve discrimination of structural substrates compared with amplitude-based thresholds alone. Prior studies have shown that LV EGM frequency analysis improves the identification of CT-defined fibrosis relative to standard voltage criteria and that systematic integration of advanced imaging into VT ablation can meaningfully alter procedural strategy, including additional mapping, epicardial access, and lesion delivery.^23,24 Consensus guidelines now provide a class IIa recommendation for preprocedural CMR in selected patients to reduce VT recurrence.¹ However, the uptake of routine LGE-CMR remains constrained by scanner availability, cost, device-related artifacts, renal function, contrast hypersensitivity, and the logistical complexity of incorporating the MRI information into the procedural workflow. EGM2Scar-AI provides increased access to this information at the point of care by extracting “imaging-like” information directly from signals already acquired in standard procedural workflows. This offers the potential for real-time, point-by-point characterization of local substrate without additional hardware or contrast administration. In this analysis, the decision threshold for scar vs. no scar was set to the thresholds corresponding to the Youden index in the development cohort.²² Practically, however, the threshold may be tuned by the operator at the point of deployment to balance sensitivity and specificity based on procedural goals, similar to other tunable parameters in contemporary mapping modules.

A key strength of this work is the creation of multimodal, spatially resolved ground truth labels used for training. We generated patient-specific 3D meshes from LGE-CMR, segmented fibrosis across the myocardial wall, and carefully coregistered these meshes to high-density electroanatomic maps. This allowed each EGM to be labeled not only as scar or no scar, but according to its location between the endocardium and the epicardium. However, prior studies have shown general, but not perfect, agreement between CMR and voltage-defined scar, and our work did not explicitly re-validate those relationships in a larger cohort.²⁵ The self-supervised model was pretrained using reconstruction tasks, followed by fine-tuning for scar classification, leveraging large volumes of unlabeled EGM data. Compared with conventional CNNs, this approach markedly improved sensitivity for scar detection while preserving specificity, particularly for mid-myocardial fibrosis, which is traditionally challenging to detect during mapping.

Despite these strengths, our findings should be interpreted in the context of several important methodologic limitations and areas for further development. First, although >130,000 EGM points were analyzed, they were derived from a relatively modest number of patients at a single center. Our study design enables the required detailed mesh creation and careful coregistration but limits the diversity of underlying pathologies, mapping strategies, and imaging protocols. The predominance of non-scar points introduced substantial class imbalance, which we mitigated through weighted loss functions but could still bias performance estimates. While point-wise analysis yielded narrow confidence intervals, patient-level subsampling revealed expectedly greater uncertainty. Epicardial scar classification demonstrated lower discrimination than endocardial and mid-myocardial scar classification but was comparable to prior studies using voltage as a surrogate for epicardial fibrosis.²⁶ While the current study supports the feasibility of the model, larger prospective cohorts are required to ensure generalizability to a wider range of patient conditions, anatomies, and to define clinically actionable thresholds. The workflows developed during this study provide the baseline for multicenter data curation.

Several modeling assumptions were necessary to make the problem tractable. We treated EGM points as independent samples, projected each point to the nearest endocardial surface vertex, and modeled the myocardial surfaces as best-fit polynomial curves. These simplifications ignore local conduction dynamics, fiber architecture, and the potential influence of catheter orientation. Geometric assumptions such as scar detection radius, conversion of scar voxel to volumetric mesh, and equal transmural partitioning may further introduce imprecision and bias to the creation of fibrosis labels used for training, especially in areas of high spatial complexity such as the LVOT and papillary muscles. However, best practices from recent literature and physiologically plausible parameters were chosen, whenever applicable, to preserve generalizability of our model. In addition, to prevent amplitude from dominating the learned representation, we scaled each channel to an absolute peak of 1, which removes direct use of peak-to-peak voltage. This choice was made to support unbiased cross-channel learning of the reconstruction model and its performance supports the notion that morphology, timing, and inter-channel relationships carry rich structural information, however the independent contribution of voltage amplitude remains untested. This design choice may reduce the model’s ability to capture certain amplitude-driven features and complicates direct comparison with standard voltage thresholds.

The findings of this study and the dataset curation pipeline provide the basis for an expanded resource including the development of a fully automated segmentation and coregistration pipeline using whole-heart models, multi-institutional datasets, and prospective workflow aligned studies. Such datasets not only provide a pragmatic tool for ablation guidance, but a patient-specific modeling framework. Integration of the transformer-derived scar maps with computational models of anatomy, mechanics, and electrophysiology, could enable ‘digital twin’ representations that are informed by both mapping and imaging.^27,28

Conclusion

EGM2Scar-AI leverages routine intracardiac electrograms to identify endocardial, mid-myocardial, and epicardial scar using a transformer-based self-supervised model benchmarked against LGE-CMR. By extracting structural information directly from contact EGMs, this approach has the potential to generate MRI-like fibrosis maps during EP procedures without the need for external imaging, and to serve as a building block for future patient-specific digital twin models in both biomedical research and precision arrhythmia care.

Supplementary Material

Supplemental Methods

Tables S1

Figures S1–S6

References 29–30

What is Known

• Characterizing deep myocardial fibrosis during ventricular arrhythmia ablation remains challenging, as standard electroanatomic mapping primarily analyzes the endocardial surface using voltage criteria.

• Although cardiac magnetic resonance imaging is the gold standard for identifying complex scar substrates, its routine intraprocedural integration is limited by accessibility, cost, device artifacts and complex workflow.

What the Study Adds

• EGM2Scar-AI, a novel transformer-based model, accurately predicts the presence and transmural distribution of endocardial, mid-myocardial, and epicardial fibrosis using routine endocardial contact and surface electrograms.

• The study demonstrates that spatial and temporal relationships within electrograms encode deep structural information, enabling real-time, three-dimensional scar characterization. during endocardial ventricular contact mapping.

• A transformer-based self-supervised model pre-trained on interpreting relationships between endocardial unipolar, bipolar, and surface electrogram signals outperforms conventional convolutional neural network models using raw waveforms or waveform-derived spectrograms.

• This approach offers a practical, point-of-care solution to identify ablation substrates without requiring external imaging, laying the groundwork for improved intraprocedural guidance and patient-specific modeling.

Nonstandard Abbreviations and Acronyms

AUPRC

area under the precision-recall curve

AUROC

area under the receiver-operating characteristic curve

CAD

coronary artery disease

EAM

electroanatomic map

ECG

electrocardiogram

EGM

cardiac contact electrogram

HF

heart failure

HTN

hypertension

ICD

Implantable cardioverter-defibrillator

LGE-CMR

late gadolinium-enhanced cardiac magnetic resonance imaging

NPV

negative predictive value

PPV

positive predictive value

STFT

short-time Fourier transform

VA

ventricular arrhythmia