Back to Projects

Choosing who not to die: AI-enhanced risk stratification for implanted defibrillators

Project ID: 2019_D19

1st supervisor: Martin Bishop, King’s College London
2nd supervisor: Pablo Lamata, King’s College London

Clinical Motivation – Reentrant ventricular arrhythmias are responsible for over 80,000 instances of sudden cardiac death (SCD) annually in the UK. Patients at risk of lethal ventricular arrhythmias often receive an implanted cardioverter defibrillator (ICD) which automatically detects arrhythmia occurrence and applies appropriate electrotherapy. However, ICDs are far from an optimal therapy, with significant clinical challenges associated with accurately assessing device need, procedural risks during the implantation (infections), and long-term complications associated with multiple electrotherapies, not to mention the adverse psychological effects of inappropriate shocks (~10% per year). Unfortunately, the majority of patients receiving an ICD suffer the significant risks and expense whilst never making use of their device, with only 2-5% having an appropriate therapy per year. More importantly, the majority of SCD events occur in patients without a device, because their elevated risk was not correctly identified. More sensitive and specific clinical risk stratification biomarkers are therefore essential to improve patient selection for device implantation.

State-of-the-art – Susceptibility to cardiac arrhythmias represents a highly complex system with intrinsic dependence on pathological structural and functional cardiac properties. Current guidelines suggest the use of a compromised LV ejection fraction (LVEF) to guide ICD implantation; however, only a third of SCD cases exhibit low LVEF. Functional biomarkers (from ECG recordings) and structural biomarkers (from MR imaging) have been independently correlated with arrhythmia risk, however, they have yet to demonstrate sufficient power to alter current clinical practice. Our detailed mechanistic modelling has suggested that path-finding of potential reentrant circuits around regions of scar will correlate closely with susceptibility to reentry. We have also shown that biophysically-detailed whole-torso models can be used to identify key arrhythmic risk biomarkers such as subtle changes in the ECG during specially designed protocols.

Key hypothesis – A risk prediction model that combines both structural and functional features, augmented by structure-function simulation-based features, will significantly enhance risk stratification for ICD requirement.

Specific Goal – To utilise rapid simulations within imaged-based models to identify susceptible reentrant arrhythmia pathways based on structure-function interactions; these will be combined with modelling-driven key ECG functional features and standard LGE derived structural features and integrated into an enhanced machine learning (ML) based risk prediction model.

Proposed Approach – An existing cohort of ~120 patients with full imaging and electrical data, as well as ICD follow-up will be used. Standard structural biomarker features will be extracted from the imaging data. Rapid path-finding simulations will be developed and performed on segmented MR imaging data, in order to locate and grade structure-function susceptibility features. Whole torso simulations will also be conducted to identify functional biomarkers to be extracted from electrical ECG data, along with standard ECG biomarkers. Structural (image-derived), functional (ECG-derived) and structure-function (simulation derived) features will be individually correlated with arrhythmia risk, and combined into a ML-based risk prediction model. The model will be trained on the existing cohort, and validated on a similar separate CRT data-base cohort.

[1] Kotu et al (2015). Cardiac magnetic resonance image-based classification of the risk of arrhythmias in post-myocardial infarction patients. Artificial Intelligence in Medicine, 64(3), 205–215.
[2] Lyon et al (2018) Computational techniques for ECG analysis and interpretation in light of their contribution to medical advances. J R Soc Interface, 15 20170821
[3] Neic et al (2017). Efficient computation of electrograms and ECGs in human whole heart simulations using a reaction-eikonal model. Journal of Computational Physics, 346, 191–211.

Back to Projects