Sustained ventricular tachycardia (VT) is a cardiac arrhythmia that originates whenever there is tissue substrate constituting a re-entrant circuit, i.e., the presence of viable myocytes (‘channels’) embedded in myocardial scar, such as post-acute myocardial infarction (AMI) scar. AMI is one of the most prevalent human diseases, and sustained VT is a major cause for sudden cardiac death (SCD) after AMI.

VT substrate ablation is an effective, invasive treatment for patients with scar-related VTs, but the rate of recurrences remains still high. It consists of introducing catheters, through a peripheral vein or artery, inside the heart. These catheters have sensors that help localising their own 3D position within the patient’s heart and record local electrical signals, thus identifying (using a computed electromagnetic navigation system) the underlying arrhythmia substrate in form of electroanatomical maps (EAM). These catheters also allow to deliver radio-frequency (RF) energy to eliminate the arrhythmia substrate. Acquiring a full EAM of the heart is a challenging and time-consuming task that may last for >3 hours, thus increasing the likelihood of procedure-related complications.

Machine learning (ML) is an emerging discipline that defines computational and statistical methods allowing computers to learn how to perform tasks based on existing data. ML has proven to be at least as effective as experts when performing heterogeneous tasks. Moreover, given that the collection of massive and heterogeneous data in clinical practice is growing, tasks in which a sufficient amount of data is available will benefit from ML. This is especially the case of image and signal processing in cardiology with the acquisition of cardiac images, electrocardiogram (ECG) recordings and other heterogeneous data.

It is possible to identify the VT site of origin (SOO) by analysing the 12-lead ECG with good accuracy. However, it is done based on visual inspection of the ECG tracings, combined with clinical expertise. On the other hand, it has been proved that VT-SOO ablation aided by pre-procedural late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) imaging results in lower need of RF delivery and improved VT recurrence-free survival. The software used to post-process LGE-CMR data allows to characterise the substrate, identifying critical parts (‘channels’) of the VT circuit and reducing the need for acquiring an EAM. However, selection of the RF ablation targets has been based on EAM findings up to date.

This project aims to evaluate and integrate into commercial software a ML pipeline for automatically identifying the VT-SOO using non-invasive information from surface 12-lead ECG recordings. This identified SOO, fused with detected ‘channels’ in post-processed pre-procedural LGE-CMR data will allow to accurately select the RF targets, hence dramatically shortening the procedure, removing inter-observer variability and likely obtaining better clinical results.