Artificial Intelligence-Enhanced Risk Stratification for Implanted Defibrillators in Ischaemic Cardiomyopathy Patients

Student thesis: Doctoral ThesisDoctor of Philosophy


In the UK, ventricular arrhythmias are responsible for over 80,000 Sudden Cardiac Deaths (SCDs) annually. Patients at risk of lethal ventricular arrhythmias often receive an Implanted Cardioverter Defibrillator (ICD); however, the current clinical guideline of a left ventricular ejection fraction (LVEF) <35% is outdated, with well-documented limitations, which has resulted in most SCD events occurring in patients without an ICD.

Late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) can characterise the phenotypic substrate for SCD non-invasively. Furthermore, advanced analysis considering changes in ventricular shape, the scar's microstructure, and its ability to provide a pro-arrhythmic substrate may shed new insight for enhanced risk prediction of SCD.

Our primary aim was to assess the utility of core infarct scar and peri-infarct zone (PIZ) characterisation by LGE CMR, building a patient-personalised composite of scar patterns to predict SCD. We further aimed to combine LV structural features to quantify the remodelling of the LV into an LV arrhythmic shape score (LVAS). Finally, we aimed to assess the structure- function vulnerability of the patient's scar substrate to sustain arrhythmias by detecting ventricular tachycardia (VT) circuits via a novel virtual induction and treatment of arrhythmias (VITA) protocol.

LGE CMR images of 437 stable ischaemic cardiomyopathy (ICM) patients and 10 years of follow-up data were analysed with bespoke image-processing tools to extract scar microstructure features and used to develop the analysis framework. 3D LV shape models were developed using Principle Component Analysis and Cox-Lasso to characterise the average arrhythmogenic LV shape and generate the patient's LVAS scores.

We developed bi-ventricular models with tetrahedral finite element meshes to apply VITA. We applied clinically preferred Cox Proportional Hazard methods alongside a machine learning (ML) approach to compare against the current clinical benchmark of LVEF and NYHA.

Both core scar and PIZ were significantly associated with the primary outcome, hazard ratio [HR] (95% CI): 1.07 (1.02-1.12), P=0.002 and HR: 1.03 (1.01-1.05), P=0.01 and the addition of PIZ and core infarct improved discrimination of the model C-Statistic from 0.64 to 0.79.

Analysing scar microstructure, PIZ entropy, PIZ components and core interface area were significantly associated with arrhythmic risk. Combining into ML models, scar patterns provided a 0.17 increase in performance compared to clinical guidelines: AUROC: 0.81 (0.81-0.82) vs 0.64 (0.63-0.65), P=0.002.
Analysis of overall LV remodelling, independent of scar, via computational-shape analysis, we showed that the LVAS metric remained independently associated with the primary endpoint: HR: 2.1 (1.5-3.0), P<0.001. C-Statistic from 0.64 to 0.75.

Finally, when identifying simulated re-entrant circuits throughout the scar, we showed the effective simulated VT cycle-length improved discrimination: HR 1.24 (1.13 - 1.37), P<0.001; C- Statistic from 0.64 to 0.72.

LGE CMR scar microstructure metrics, LV shape remodelling scores and virtual simulation of VT over bi-ventricular heart models were used to develop an independently predictive ML model for SCD risk beyond conventional predictors used in ICD implantation guidelines. These results signify the potential for a more personalised approach in determining ICD candidacy in ICM.
Date of Award1 Mar 2024
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorMartin Bishop (Supervisor) & Pablo Lamata de la Orden (Supervisor)

Cite this