Machine learning-assisted mapping of VT ablation targets: progress and potential

Ventricular tachycardia (VT) remains a leading cause of sudden cardiac death, even in the era of implantable defibrillators and catheter ablation. Drawing on his own experimental and clinical research, Dr Michele Orini shares how machine learning can help to identify critical VT ablation targets, pointing to a safer, data-driven approach to treating life-threatening ventricular arrhythmias.

Ventricular tachycardia (VT) is a serious and potentially life-threatening heart rhythm disorder, originating in many patients from areas of scar tissue within the heart muscle. This may result from a previous heart attack or underlying cardiac conditions.

Regardless of its cause, the scar tissue disrupts the normal conduction of electrical signals, allowing abnormal loops of electrical activity to form and sustain the arrhythmia. When these rapid electrical loops occur within the ventricles, the heart cannot pump blood effectively, compromising circulation and blood pressure. If left untreated, VT can lead to sudden cardiac death.

Implantable cardioverter-defibrillators (ICDs) are designed to terminate life-threatening arrhythmias through potent electrical shocks or anti-tachycardia pacing. It is estimated that approximately 200,000 ICDs are implanted worldwide each year.

While ICDs are highly effective in preventing sudden cardiac death, patients experiencing frequent ventricular arrhythmias may face recurrent and painful shocks, which are associated with increased anxiety and reduced quality of life. In addition, repeated shocks may further impair cardiac function, particularly in patients with underlying heart disease.

Consequently, significant research has focused on treatments that prevent ventricular arrhythmias from occurring in the first place. Antiarrhythmic drugs initially offered considerable promise in the 1990s, but the balance between benefits and risks, given their side effects, has proved disappointing.

Over the past two decades, procedures that directly target the arrhythmogenic substrate have been developed. Catheter ablation of VT has become an established treatment for patients with recurrent, life-threatening arrhythmias, and trials have demonstrated its greater efficacy compared with medical therapy.¹

Standard approaches for catheter ablation and limitations

Catheter ablation is an invasive procedure used to treat arrhythmias, including VT, by targeting and eliminating the cardiac tissue responsible for abnormal electrical activity. Via the catheter, advanced mapping systems then collect tens of thousands of intracardiac signals to generate detailed three-dimensional maps of electrophysiological properties relevant to arrhythmogenesis.

Typically, ventricular arrhythmia is then induced through programmed electrical stimulation, allowing expert clinicians to identify regions that generate and sustain the arrhythmia. Once identified, this tissue is typically destroyed using radiofrequency energy, interrupting abnormal electrical circuits while preserving normal heart function.

Despite considerable advances, current strategies for catheter ablation of VT have several limitations. The procedure is lengthy, typically lasting three to five hours. Inducing VT during the electrophysiological study allows direct mapping of the electrical circuits responsible for the arrhythmia, but it carries a non-negligible risk, as blood pressure may not be sustained, often requiring cardioversion or mechanical circulatory support. Finally, VT recurrence rates after the procedure remain high, at approximately 30–50%.

An emerging alternative strategy aims to identify accurate ablation targets – the cardiac sites critical for VT initiation and maintenance – while the patient remains in sinus rhythm. Although this approach improves procedural safety, it makes identifying VT-critical sites more challenging, as the arrhythmia circuit cannot be directly mapped and must instead be inferred. This complex task is currently performed by highly skilled clinicians relying on extensive expertise and experience.

The role of machine learning in supporting VT care

Predicting ablation targets from cardiac mapping data acquired in the catheterisation laboratory is a problem well suited to novel computational approaches. Accordingly, ongoing research is leveraging machine learning and other types of artificial intelligence (AI) to automate this process, with the goals of improving speed, accuracy and scalability.

AI offers considerable potential in this context by enabling the solution of complex classification problems and the identification of patterns within large datasets. While AI has demonstrated substantial promise in several areas of electrocardiography,² as well as in improving procedures such as catheter ablation for atrial fibrillation,³ its application in identifying VT ablation targets remains extremely limited.

In our recent study, published in the European Heart Journal – Digital Health,⁴ we used data from a porcine myocardial infarction model to investigate whether machine learning could help identify potential ablation targets. A key advantage of this large-animal model is that VT circuits can be mapped over longer durations and in greater detail, while allowing the simultaneous evaluation of multiple mapping strategies.

The experiments were conducted in Copenhagen, Denmark, using the latest-generation cardiac mapping catheters to characterise the electrophysiological properties of myocardial tissue surrounding the infarct scar during sinus rhythm and under programmed electrical stimulation.

VT was subsequently induced, and the critical components of the VT circuit were identified using a state-of-the-art cardiac mapping system of the same type routinely used in clinical practice.

We applied advanced signal-processing techniques to extract electrophysiological features from 35,068 intracardiac electrograms recorded during sinus rhythm and programmed electrical stimulation.

Multiple machine learning models – including k-nearest neighbours, random forests and support vector machines – were evaluated for their ability to predict the location of critical VT circuit sites. These approaches proved highly effective in identifying VT-critical regions using data acquired during sinus rhythm or pacing, without the need to induce VT during the procedure.

Importantly, our study is the first to demonstrate that machine learning methods leveraging features derived from intracardiac electrograms can support clinicians in localising VT ablation targets using substrate mapping.

Building on these findings, we subsequently conducted a study in patients undergoing VT ablation, using clinical data acquired during procedures in which VT was induced and successfully mapped. As in our previous work, machine learning models were trained exclusively on data collected during sinus rhythm or pacing to indirectly infer the location of critical VT sites.

Preliminary, unpublished results are encouraging and highlight the potential of a more advanced approach based on graph neural networks (GNNs). Unlike conventional models, GNNs learn not only from the morphology of intracardiac signals but also from their spatial relationships. This architecture appears particularly well suited to cardiac mapping applications, as adjacent cardiac sites are physically connected and spatial heterogeneity in electrical activation and repolarisation is a key determinant of arrhythmogenesis.

Translating VT ablation research into clinical practice

These findings are highly encouraging for a procedure that, when successful, can be life changing and, in many cases, lifesaving. In parallel, several complementary technologies have emerged in recent years, with the potential to further improve VT ablation outcomes. These include advanced medical imaging techniques that better characterise arrhythmogenic substrates, as well as digital twin approaches capable of reproducing electrophysiological mechanisms that cannot be directly measured in vivo.

Nevertheless, meaningful clinical translation will require substantial further effort. The value of these novel approaches, including the methods proposed in our work, must be validated in large, randomised clinical trials. The corresponding algorithms will also need to be robustly integrated into clinical mapping systems suitable for routine use during electrophysiological procedures.

Together, these developments point toward a future in which VT ablation is guided by data-driven, patient-specific models that improve accuracy and safety while reducing procedural duration and cost.

If successfully translated into clinical practice, such approaches could fundamentally transform the management of ventricular arrhythmias.

Author

Michele Orini MEng PhD
Associate professor in healthcare engineering, Department of Biomedical Engineering, King’s College London, UK

References

1. Sapp JL et al. Catheter Ablation or Antiarrhythmic Drugs for Ventricular Tachycardia. N Engl J Med 2025;392:737–47.
2. Siontis KC et al. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Rev Cardiol 2021;18:465–78.
3. Deisenhofer I et al. Artificial intelligence for individualized treatment of persistent atrial fibrillation: a randomized controlled trial. Nat Med 2025;31:1286–93.
4. Wang X et al. Machine learning approach for automated localization of ventricular tachycardia ablation targets from substrate maps: development and validation in a porcine model. Eur Heart J Digit Health 2025;6:645–55.