CMR Radiomics and AI-ECG Predict Ventricular Arrhythmias and Sudden Cardiac Arrest in Ischemic Heart Disease

Background: Many patients with left ventricular ejection fraction (LVEF) > 35% and coronary artery disease (CAD) are at risk for sudden cardiac arrest (SCA) but are not included in current primary prevention guidelines for implantable cardioverter defibrillators (ICDs). Artificial intelligence applied to 12-lead ECGs (AI-ECG) is of interest in these patients for prognostication, but there are gaps in our approaches to implementation and also in our understanding of how AI-ECG parameters are related to LV structural and functional findings from CMR, particularly scar characteristics and strain. AI-ECG and CMR could also provide synergistic prognostication for ventricular tachycardia/ventricular fibrillation (VT/VF) and SCA.

Methods: We performed simultaneous AI-ECG and CMR assessments of scar and strain in patients with CAD and left ventricular (LV) dysfunction. Scar burden was determined from late gadolinium enhancement (LGE) (5SD), and scar features were assessed with LGE radiomics (LGE-RAD). A deep learning model (LaMoD) assessed global and regional strain (ref. 1). AI-ECG was assessed using both the widely used 1-D convolutional neural network approach (CNN) (ref. 2) and also functional principal component decomposition (FPCD) of ECG waveforms in 12 leads (ref. 3), which has the advantage of interpretability based on the ECG leads predictive of the outcome.

Results: Among 455 patients with CAD and LV ejection fraction (LVEF) ranging from 4-50%, 232 patients had LVEF > 35%, 223 had LVEF ≤ 35%, and 59 patients had the composite outcome of sustained VT/VF or SCA. In patients with LVEF > 35%, LGE burden predicted this outcome with an AUC of 0.73 (95% CI 0.57 to 0.84), but LGE-RAD features had a higher AUC of 0.84 (95% CI 0.76 to 0.91) (p< 0.05) (Figure 1), indicating better prediction with scar features vs. scar burden. The combination of AI-ECG and CMR with LGE-RAD had the highest AUC (0.94, 95% CI 0.89 to 0.97), demonstrating synergistic prediction. LGE-RAD principal component 3 (PC3) was associated with spatial intensity variations among neighboring voxels and had the greatest association with the outcome. Patients with LGE-RAD PC3 > median (versus ≤ median) had different CNN latent feature profiles (p=0.0035) and greater values of the ECG lead V2 FPCD score (p=0.029) (Figure 2). With respect to CMR strain, ECG FPCD scores in lead V1 were decreased with global peak circumferential strain (Ecc) > median (p=0.00035) and peak anteroseptal Ecc > median (p=0.035), while global Ecc > median was associated with a different ECG CNN latent feature profile (p=0.0048) (Figure 3).

Conclusion: Radiomics of CMR improves prediction beyond standard LGE for sustained VT/VF and SCA. Radiomics characterization of variability of pixel intensity within scar and parameters from global/regional strain offers explainability of AI-ECG findings and has independent prognostic value. This approach could have high impact for patients with LVEF > 35% who currently do not have a primary prevention indication for ICDs.

Figure1: Prediction of VT/VF and SCA with CMR and AI-ECG.
Figure 2: LGE Radiomics for Structural Correlates of AI-ECG Parameters.
Figure 3: Deep Learning of Strain from Cine Images for Structural Correlates of AI-ECG.