Premature ventricular contraction–mediated ventricular fibrillation: Clinical characteristics, application of machine-learning algorithm, and outcomes of catheter ablation—A case series

Abstract

Background

Although premature ventricular contractions (PVCs) are commonly seen, in some patients, PVCs are associated with malignant ventricular arrhythmias including ventricular fibrillation (VF).

Objective

This study aimed to assess electrocardiographic (ECG) characteristics and apply a machine-learning (ML) algorithm in patients with PVC-triggered VF.

Methods

We analyzed data from an international cohort of 82 patients undergoing ablation for PVCs (41 with PVC-triggered VF and 41 controls). We evaluated the prevalence of ECG characteristics in patients with PVC-triggered VF, including (1) early repolarization (ER) in inferior/lateral leads and (2) QRS notching of the sinus beat or PVC, and also applied an ML algorithm to assess the differences in the 2 patient cohorts.

Results

In 41 patients with PVC-triggered VF, there was a median of 8 implantable cardioverter-defibrillator shocks per patient before PVC ablation. The mean coupling interval of the PVC to the antecedent sinus beat was 313 ± 130 ms. Compared with controls, ER (39% vs 17%) and QRS notching (71% vs 24%) were significantly more prevalent in the PVC-triggered VF group. After a median of 1 ablation (interquartile range 1–3), 82% of patients remained free of ventricular tachycardia/VF and implantable cardioverter-defibrillator shocks over a median follow-up of 400 days (90–2490). The ML ECG algorithm demonstrated reasonable discrimination between the 2 groups (area under the receiver-operating characteristic curve 0.85 [0.56–1.0]). Anterior ST-segment deviation and left bundle branch–like delay were salient contributors to ML prediction.

Conclusion

In patients with PVC-triggered VF, ER and QRS notching were more prevalent than in patients with PVC without VF. An ML-based ECG algorithm effectively distinguished between the 2 groups.

Key Findings.

• ▪Our study analyzed an international multicenter cohort of 82 patients who underwent catheter ablation for premature ventricular contractions (PVCs) (PVC-triggered ventricular fibrillation [VF] n = 41; control group of patients without ventricular tachycardia [VT]/VF history n = 41).
• ▪The patients with PVC-triggered VF had a significantly higher prevalence of 12-lead electrocardiographic (ECG) abnormalities, such as (1) notching of QRS complexes and (2) early repolarization of either inferior or lateral leads (39% vs 17% in the controls; P = .04).
• ▪When we applied a machine-learning (ML) ECG algorithm (FactorECG), we observed that it had a reasonable level of discrimination between the 2 groups of patients (area under the receiver-operating characteristic curve 0.85).
• ▪Explainability analyses identified anterior ST-segment deviation and a left bundle branch–like delay of the ECG waveform as salient contributors to the ML-based prediction model.
• ▪After a median of 1 ablation procedure per patient, 82% of patients remained free of sustained VT/VF and implantable cardioverter-defibrillator shocks over a median follow-up period of 400 days.

Introduction

Although the occurrence of premature ventricular contractions (PVCs) is relatively common¹ and often considered benign in otherwise healthy individuals without structural heart disease,² there is an emerging recognition of malignant risk in a subset of patients who develop PVC-induced ventricular fibrillation (VF). For example, in the acute and subacute setting after myocardial infarction, PVCs can trigger VF and catheter ablation of these triggering PVCs can be an effective tool to mitigate the risk of ventricular arrhythmias.^3,4 PVC-triggered VF has now been recognized in the absence of structural heart disease and underlies the diagnosis of idiopathic VF.⁵ Short-coupled PVCs and those originating from the His-Purkinje system have been identified as triggers for VF in patients with previous myocardial infarction,⁶ and their role has also been described in patients with inferolateral J-wave syndrome.^7,8 These findings underscore the hypothesis that the mechanism of VF, more broadly, may be an interaction between vulnerable substrate and triggering PVC. Previous work in patients with idiopathic VF has generally centered on the triggering PVC including modes of initiation and coupling intervals between PVC and sinus rhythm.⁹ However, presently lacking are clinically actionable strategies to identify vulnerable substrates and the risk of PVC-induced VF in individuals with PVCs.

Previous studies in the general population have suggested that some 12-lead electrocardiographic (ECG) characteristics such as early repolarization (ER) and QRS notching may be generally associated with a higher risk of SCD.^10,11 Recent work has demonstrated that deep learning analysis of unstructured ECG waveforms may identify salient signatures of incident cardiovascular risk, including, for example, the risk of PVC-induced cardiomyopathy.¹² Whether such a deep learning approach could identify salient substrate in those with PVC-triggered VF is unknown but could provide a novel diagnostic tool to inform risk and management strategies.

Therefore, in this study, we evaluated the incidence of ECG abnormalities (ER pattern, QRS notching) in patients with PVC-triggered VF. We examined the predictive accuracy of a deep learning ECG algorithm (FactorECG) for PVC-triggered VF compared with PVC without a history of VF. Finally, we performed explainability analyses of the deep learning algorithm to identify which components of the ECG are most related to the prediction of VF in individuals with PVCs.

Methods

Study population

The retrospective cohort included patients from an international (United States and Germany) multicenter registry of individuals undergoing catheter ablation for PVC-triggered VF between November 2018 and September 2024. The study was approved by the institutional review board (University of Toledo, LOI: P25-2025 in July 2025) and all the participating centers. The study conformed to the principles of the Declaration of Helsinki and received approval from the human research protection and ethics committee. The requirement for an informed consent was waived owing to the retrospective nature of the study and the use of deidentified data. The ECGs were deidentified for analyses of the deep learning (FactorECG algorithm). The first author (A.M.) takes responsibility for its integrity and data analysis.

A total of 82 patients were included in the analytical cohort (41 individuals with PVC-triggered VF and 41 controls). We excluded patients with a diagnosis of ER syndrome and Brugada syndrome and those with polymorphic ventricular tachycardia degenerating into VF. In addition, we excluded patients in whom VF was noted in the setting of active myocardial ischemia. The control group comprised patients (n = 41) undergoing catheter ablation for symptomatic PVC and without a history of sustained ventricular arrhythmias during the study period.

ECG analysis

We analyzed data from the baseline preablation ECG (either in sinus rhythm or with clinical PVCs). To assess the presence of repolarization abnormalities and notching of the QRS complexes in sinus rhythm or that of PVCs, the ECGs were independently assessed by both A.M. and E.S., and both reviewers were blinded to clinical characteristics including classification of PVC with or without VF. Discordance in assessment was resolved by an independent third reviewer (M.G.).

Based on previously published literature, notching of the QRS complex may represent heterogeneity in electrical conduction and regional conduction delay^13,14 and could also be a sensitive marker for myocardial scar.¹⁵ It has also been postulated that subtle disruptions in depolarization are best detected in lateral leads and more prominent disruptions are best seen in inferior leads. Therefore, we focused on the assessment of QRS notching and ER patterns in these leads. On the 12-lead ECG recorded at 25 mm/s, we assessed the following abnormalities in both patients.

• 1)ER pattern in at least 2 contiguous leads, either the inferior or lateral leads. ER was characterized as J-point elevation in keeping with the guidelines of the consensus statement.¹⁶
• 2)Notching or fragmentation of the QRS complex of either the native sinus beat or the PVC complex¹¹
• 3)For the assessment of notching or QRS fragmentation, we adjudicated it to be a positive finding only if it was present in 2 contiguous leads.
• 4)Coupling interval of PVCs: this was defined as the time interval between the end of the QRS complex of the preceding sinus beat and the onset of the PVC.
• 5)Duration of QRS interval of PVCs: the interval measured from the beginning to the end of the QRS interval

Machine-learning algorithm based on an explainable deep-neural network

Deep learning ECG analysis was performed using FactorECG, a variational autoencoder, the derivation of which has been previously published.¹⁷ The original FactorECG model was trained on 1.1 million ECGs and used a variational autoencoder consisting of 3 parts: an encoder; the FactorECG, which consisted of 32 continuous factors; and a decoder, which reconstructed ECGs using these 32 factors. These 32 factors are interpretable and represent physiological processes (such as atrial and ventricular depolarization, conduction properties). After training, the encoder can be used to convert any median beat ECG into the FactorECG format.^17,18 The FactorECG algorithm can be accessed using the hyperlink: https://pln.ecgx.ai. As the next step, the algorithm performs factor standardization, wherein each of the 32 latent factors is normalized to z scores. The purpose of z-score normalization is to ensure that all these factors are on the same scale that allows for direct comparison of effect sizes. For the purpose of our study, we used a comparison framework of these features between the 2 patient cohorts. This was also different from the traditional use of machine-learning (ML) techniques, given that we used a hybrid approach that combined the use of ML and traditional statistical methods. An example of how odds ratios were generated using the FactorECG algorithm as part of a hybrid-ML method is presented in the Supplemental Material.

Feature importance

In the context of the applicability of FactorECG, feature importance reflects the relative contribution of each extracted ECG factor to the output of the predictive model. Feature importance is also different from end-to-end ML models, given that it does not refer to a learned weight from a neural network. In the original FactorECG algorithm, feature importance was quantified using the absolute standardized coefficients (β) derived from the multivariable prediction model after normalization using z scores of all the ECG factors. It reflects the relative contribution of each extracted ECG factor (ie, factors that represent ventricular depolarization and repolarization) to the overall predictive model’s output. We acknowledge that although a higher value of feature importance indicates a stronger influence on model prediction, it does not imply causality or a necessary physiological association.

Preablation imaging

All patients who underwent PVC ablation had either an echocardiogram or a cardiac magnetic resonance imaging at baseline. The decision and selection of imaging modality before PVC ablation were driven by the operator’s discretion.

Ablation procedure

We also collected electroanatomic mapping data from all patients who underwent ablation for the clinical PVC, which was concluded to be the trigger for PVC-VF. During ablation, the PVC that triggers VF was targeted for ablation as the clinically relevant PVC. For induction of PVCs, a combination of burst pacing and isoproterenol infusion was used. Activation mapping was performed if PVCs were present. If a triggering PVC could not be induced, then pace mapping was used to guide catheter ablation. If elimination of the clinical PVC that triggers VF was not possible, then we adopted a substrate modification approach for denetworking of the Purkinje network as the primary ablation strategy.¹⁹

Follow-up

Follow-up data on clinical outcomes and periprocedural complications were collected. Specific data on the recurrence of ventricular arrhythmias were collected from a combination of ECGs, interrogations, and remote transmissions from implantable cardioverter-defibrillators on the follow-up clinic visits.

Statistical analysis

We compared continuous and categorical variables between the 2 groups of patients using the Student t test and the Fisher exact test, respectively. The prevalence of ECG abnormalities (ER pattern and notching of QRS) and the location of PVC ablation sites during ablation are presented as percentages. For the assessment of 32 factors of FactorECG and which factors encoded the most relevant information, we used logistic regression models. To assess the performance of FactorECG as a diagnostic model to discriminate between PVC-triggered VF and PVC without VF, we used the receiver-operating characteristic curve to calculate the concordance statistic. All the analyses were performed using Stata software (StataCorp LLC, College Station, TX).

Results

In our multicenter study, a total of 41 patients underwent catheter ablation for PVC-triggered VF. The median number of implantable cardioverter-defibrillator shocks received before catheter ablation was 8 (interquartile range [IQR] 2–60) per patient.

Approximately one-third of patients (n = 14 [34%]) presented with electrical storm before undergoing catheter ablation. On the preablation 12-lead ECG, most patients with PVC-triggered VF (30 of 41 patients [73%]) had a right bundle branch block PVC morphology. Compared with control patients, the coupling interval of the PVC to the antecedent sinus beat was significantly shorter in the PVC-triggered VF group (313 ± 130 vs 436 ± 57 ms; P < .001). The summary of the comparison of baseline characteristics between the 2 groups of patients is presented in Table 1.

Table 1.

Comparison of baseline characteristics of patients who had PVC-triggered VF

Clinical characteristics	PVC-VF subgroup (n = 41)	Control subgroup (n = 41)
Age at the time of most recent ablation (y)	49 ± 16	55 ± 12
Male gender	26 (63.5%)	18 (44%)
Left ventricular ejection fraction (%)	48 ± 12	51 ± 8
Coronary artery disease	14 (34%)	11 (27%)
Previous myocarditis	5 (12.2%)	3 (7%)
Scar/LGE on preablation cardiac MRI	10 (24.3%)	3 (7%)
Patients who experienced any ICD shock before ablation	30 (73%)	NA
Median number of ICD shocks per patient	8	NA
Patients presenting with electrical storm before ablation	14 (34%)	NA
Average number of ablations per patient	1.5 ± 0.7	1.3 ± 0.6
Follow-up duration (d)	585 ± 519	178 ± 106

ICD = implantable cardioverter-defibrillator; LGE = late gadolinium enhancement; MRI = magnetic resonance imaging; PVC = premature ventricular contraction; VF = ventricular fibrillation.

We next evaluated the prevalence of 12-lead ECG abnormalities in those with PVC-triggered VF compared with individuals with PVC without VF. We found that 39% of individuals with PVC-triggered VF had an ER pattern in the inferior of lateral leads compared with 17% of individuals with PVC without VF (P = .04). Similarly, individuals with PVC-triggered VF had a higher prevalence of QRS notching than control patients (71% vs 24%; P < .001) (Figure 1).

Figure 1.

Notching of the sinus rhythm QRS complex (red arrow) and notching of the premature ventricular contraction complex (blue arrow) in a patient presenting with premature ventricular contraction–triggered ventricular fibrillation.

Ablation findings and clinical outcomes

During catheter ablation in patients with PVC-triggered VF, the median number of PVCs that were targeted was 1 (IQR 0–4). An ablation approach aimed at either targeting or denetworking of the Purkinje potentials was performed in two-thirds of these patients (61%) (Figures 2 and 3). The predominant location of PVCs in the PVC-VF group of patients was in the left ventricular septum (61%), followed by the left fascicular system and papillary muscles (12% each) (Figure 4). Noninducibility of PVC-triggered VF as an acute procedural endpoint was achieved in all patients. The results of ECG features and ablation findings are presented in Table 2. After a median of 1 ablation procedure (IQR 1–3), 83% of the patients (34 of 41) remained free of VF over a median follow-up period of 388 days (IQR 42–2490).

Figure 2.

PVC-induced VF and ablation targets (Purkinje potentials) during PVC-triggered VF ablation. PVC = premature ventricular contraction; VF = ventricular fibrillation.

Figure 3.

Ablation lesions and left anterior and posterior fascicles.

Figure 4.

Location of ablation for PVC targeted in patients with PVC-triggered VF. LV = left ventricular; PVC = premature ventricular contraction; RVOT = right ventricular outflow tract; VF = ventricular fibrillation.

Table 2.

Ablation characteristics and clinical outcomes after ablation for PVC-triggered VF

ECG characteristics	PVC-VF subgroup (n = 41)	Control subgroup (n = 41)	P value
QRS duration of PVCs (ms)	141 ± 28	145 ± 11.1	.40
Coupling interval of PVCs (ms)	313 ± 130	436 ± 57	.001
Left bundle branch block PVC morphology	5 (12%)	22 (54%)	.002
Right bundle branch block PVC morphology	30 (73%)	19 (46%)	.01
Notching of the PVC QRS complex	29 (71%)	10 (24%)	.001
Early repolarization pattern	16 (39%)	7 (17%)	.04
Ablation characteristics
Number of PVCs observed during ablation	1.4 ± 0.9	1.4 ± 0.6	.95
VF induced during ablation	28 (68.3%)	0	NA
Purkinje potentials targeted during ablation	25 (60.9%)	0	NA

ECG = electrocardiographic; NA = not available; PVC = premature ventricular contraction; VF = ventricular fibrillation.

Deep learning ECG discrimination of PVC-triggered VF

Using an autoencoder deep learning ECG algorithm, we identified 8 key ECG “factors” associated with the prediction of PVC-triggered VF compared with PVC without VF. When combining these key factors, the deep learning ECG algorithm demonstrated a reasonable discrimination between PVC-triggered VF and PVC without a history of ventricular arrhythmia (area under the receiver-operating characteristic curve 0.85 [0.56–1.0]). In explainability analysis, the 2 most salient ECG signatures associated with PVC-triggered VF were (1) anterior ST-segment deviation and (2) left bundle branch–like morphology (Figure 5). In Figure 5B, the relative values of “feature importance” represent the contribution of each deep learning–derived ECG factor to the overall model. Although a higher value of feature importance (as shown in Figure 5B) might indicate a stronger contribution for prediction, it does not imply a causal or a necessary physiological association.

Figure 5.

Use of FactorECG in predicting PVC-induced VF. A: ROC curve analysis for FactorECG. B: Explainability of each FactorECG “factor” in the prediction of PVC-induced VF. Feature importance indicates the relative contribution of each factor to the overall model and is a unitless entity. C: Boxplot of AUROC and AUPRC. D: Visual representation of median beat ECG differences between patients with PVC-induced VF and those undergoing ablation for PVC not inducing VF. AUPRC = area under the precision-recall curve; AUROC = area under the receiver-operating characteristic curve; ECG = electrocardiographic; PVC = premature ventricular contraction; ROC = receiver-operating characteristic; VF = ventricular fibrillation.

Discussion

In this multicenter cohort of patients with PVCs with and without VF, we showed that patients with PVC-triggered VF had a higher prevalence of 12-lead ECG abnormalities, including ER pattern and QRS notching, than patients with PVCs and no history of ventricular arrhythmia. We also show that catheter ablation of trigger and substrate yielded freedom from VF in most patients after a median of 1 procedure. We demonstrated that a deep learning ECG algorithm applied to the sinus rhythm ECG effectively discriminated patients with PVC-triggered VF compared with those with PVCs and no history of VF. Specific sinus rhythm ECG signatures that were associated with the prediction of PVC-triggered VF were anterior ST-segment deviation and left bundle branch block morphology.

Although PVCs are commonly encountered in clinical practice, there is now well-established recognition that a subset of individuals with PVCs can be at risk of PVC-triggered VF and sudden death.²⁰ Although there has been substantial evolution of our understanding of interventional approaches to this condition, including PVC ablation and denetworking of the Purkinje system, effective tools for risk stratification in the broader population of patients with PVCs are lacking.²¹ Previous studies have suggested that the coupling interval between a PVC and the antecedent sinus beat could be a helpful arbiter of ventricular arrhythmia risk in those with PVCs. For example, in an observational study by Hamon et al,²² a variable coupling interval was associated with a higher risk than a fixed PVC coupling interval (either short or long). However, more recent multicenter data from the THESIS and CASPER registries have indicated risk in individuals with PVC regardless of coupling interval.^5,9 Taken together, these conflicting data underscore the limitations of relying on coupling interval as a risk stratification tool in individuals with PVCs. To that end, there has been a parallel interest in understanding the role of myocardial substrate in the pathogenesis of PVC-triggered VF²³ and the relationship between this abnormal substrate and the triggering PVC.²⁴ For example, high-density electrogram mapping in patients with idiopathic VF demonstrated frequent colocalization of abnormal electrograms and drivers of VF.²⁵ Areas of identified electrical abnormality were often subtle and regionally small, underscoring the potential limitations of other screening methods such as cardiac magnetic resonance imaging as a risk stratification tool.

Given these potential limitations, we sought to evaluate whether the 12-lead ECG—a low-cost and scalable tool—could be used to enhance risk identification in patients with PVCs. We evaluated both visually apparent ECG abnormalities and the utility of a deep learning algorithm incorporating ECG waveforms. First, we showed that visually apparent ECG abnormalities (inferior ER, QRS notching) were significantly more commonly present in patients with PVC-triggered VF than patients with PVC without VF. We hypothesized that these findings, mechanistically, are reflective of impaired electrical dynamics and repolarization heterogeneity that underlie a vulnerable substrate in patients with PVC-triggered VF. Previous studies have linked ER with epicardial electrogram abnormalities generally,²⁶ and at least 1 other group has similarly identified ER and fractionated QRS morphology in patients with PVC-triggered VF.²⁷

Second, we demonstrated the role of a deep learning algorithm to distinguish patients with PVC-triggered VF from those with PVC without VF. Previous studies have demonstrated the utility of deep learning PVC algorithms to identify patients at risk of PVC-induced cardiomyopathy, potentially identifying ECG features capturing impaired excitation-contraction coupling and autonomic dysfunction.¹² Others have also suggested that deep learning analysis of the ECG may be able to predict incident ventricular arrhythmias in patients with structural heart disease.^28,29 Here, we showed as a proof of principle that a deep learning algorithm could distinguish between patients with PVC-triggered VF and those with PVC without VF. In explainability analysis, we found that anterior ST-segment deviation and left bundle branch–like morphology were salient predictors of risk. However, on balance, given the complex pathophysiology and role of several factors that lead to VF, a multimodal fusion ML model that could incorporate imaging markers (such as late gadolinium enhancement), electroanatomic mapping, and autonomic parameters could provide a more comprehensive approach for risk stratification. We hypothesized, from a mechanistic standpoint, that these features may capture subtle depolarization abnormalities that may be salient to risk in those with PVC-triggered VF. Given the low-cost and scalable nature of the 12-lead ECG, if these findings can be replicated in a larger prospective evaluation, such a deep learning–based approach could transform the paradigm of malignant arrhythmia screening in individuals with PVCs. If the results of ML-based risk stratification are externally validated, then the patients with PVCs who are at “high risk” of VF could be considered for an early ablation approach.

Limitations

We acknowledge several limitations of our study. First, this is a retrospective cohort study from an international multicenter study group. Whether or not these findings and specifically ablation outcomes are generalizable to other centers is not known. Second, although the assessment of ECG abnormalities was performed with blinded review and discrepancies resolved by third-party assessment, we note that real-world evaluation of these abnormalities may be subject to ascertainment variability, which could influence clinical utility. Third, the interventional approach to PVC-triggered VF ablation was not prospectively harmonized, which could have introduced heterogeneity into the analysis of outcomes. That being said, by virtue of being a real-world cohort in which ablation strategy was deferred to individual operators, we believe that this work may be more generalizable because it reflects the empirical heterogeneity of ablation approach in this condition. Fourth, the sample size of our study is relatively small, which is an important limiting factor for the feasibility of statistical analyses (such as multivariable regression and time-to-event analysis). Similarly, the limited sample size was also a limiting factor that could have contributed to overfitting in the ML algorithm’s interpretation.

Finally, given the size and retrospective nature of this cohort, we would emphasize that these findings reflect key pilot data, which are exploratory and hypothesis generating rather than practice changing. Further prospective and large-scale assessment of a multimodal approach that incorporates clinical, imaging, and ML-based assessment would lend further strength to our findings. We also acknowledge that a prospective validation in a larger, unselected patient cohort will be necessary before the adoption of an ML model as a clinically actionable tool.

Conclusion

In this retrospective, international multicenter cohort study, we showed that visually assessed abnormalities (inferior repolarization abnormality, QRS notching) and deep learning–based analysis of a routine 12-lead ECG were able to effectively distinguish patients with PVC-triggered VF from those with PVC without VF.

Disclosures

A.M.: consulting for Biotronik and Mariana AI and equity with Pulse Biosciences. R.V.L.: cofounder of FactorECG. The other authors have no conflicts of interest to disclose.