Automated detection of ventricular tachycardia locations using a combined physics and deep learning approach

Student thesis: Doctoral ThesisDoctor of Philosophy


Ventricular tachycardia (VT) is a serious cardiac arrhythmia, and an important cause of sudden cardiac death (SCD) and morbidity. Therefore, the management of the tachycardia is of importance to prevent degeneration into fibrillation, SCD and improve quality of life. Implantable cardiac electronic devices are the first-line therapy for most VT patients at high risk. However, the most effective, curative option against incessant VT is considered ablation therapy. The success of an ablation procedure is heavily dependent on the accurate localisation of the sites responsible for the initiation and maintenance of the tachycardia (e.g. focal ectopy, isthmus, exit/entrance). The identification of these optimal ablation targets is carried-out via either invasive and time-consuming electrophysiological (EP) mapping strategies (e.g. pace-mapping) or non-invasive modalities, which all rely on the surface electrocardiogram (ECG). Difficulties in inducing the arrhythmia in haemodynamically unstable patients and/or with limited conventional treatment options might challenge the acquisition of the ECG, and affect correct identification of ablation targets. In addition, recurrence of the VT may occur if the clinical episode has not been properly targeted. These limitations of current ablation mapping strategies warrant the development of more specifically targeted solutions to increase ablation success rate and terminate VT in the long-term.

With this research, we aim to improve ablation planning and identification of VT ablation targets by utilising the information stored as electrograms (EGM) in implanted devices, that the majority of VT ablation patients have in-situ. By doing so, we could target the clinical episodes, as well as reduce the need for VT induction, ultimately improving safety, speed and accuracy of ablation. We intend to achieve this by using computational models to simulate implanted device EGMs and investigate their utility in in-silico pace-mapping. In addition, we will use computational simulations to generate a vast library of ECGs and EGMs of different VT episodes, which will be useful in conjunction with deep learning approaches to automate the localisation of VT critical sites. Finally, we aim to evaluate our in-silico platforms within clinical settings.

Our research was successful in showing the utility and power of simulated, multi-vector EGM recordings for the identification of critical VT sites in both in-silico pace-mapping and automated algorithms. We reported results comparable to using the ECGs. In addition, we were able to evaluate our platforms in clinical settings, when using clinical VT ECGs. Despite the lack of EGM recordings of the clinical VT episodes, given the extensive computational analyses carried-out throughout this Thesis, we believe that the performance of our platforms would return promising results when using real VT EGMs, thus underscoring the importance of this proof-of-concept research to improve VT management and ablation planning.
Date of Award1 Jan 2023
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorMartin Bishop (Supervisor) & Andrew King (Supervisor)

Cite this