Derivation and internal validation of a novel cardiac magnetic resonance risk score for prophylactic implantable-cardioverter defibrillator therapy in ischemic cardiomyopathy
A Unger, A G Pavon, J Amar, L Bergamaschi, S Toupin, M Valgimigli, C Pizzi, A I Guaricci, N Carrabba, L Fusini, A Varga-Szemes, J Schwitter, T Pezel, M Guglielmo, G PontoneAbstract
Background
Sudden cardiac death (SCD) remains a major cause of mortality in ischemic cardiomyopathy (ICM). Current guidelines recommend implantable cardioverter-defibrillator (ICD) implantation mainly based on left ventricular ejection fraction (LVEF), which has limited specificity for arrhythmic events. Cardiovascular magnetic resonance (CMR) provides reproducible markers of myocardial remodeling and scar burden, yet no validated CMR-based machine learning (ML) risk score currently exists for arrhythmic risk stratification.
Purpose
To develop and validate a CMR-based ML risk score for predicting major arrhythmic adverse cardiac events (MAACE), defined as appropriate ICD therapy, sustained ventricular arrhythmias, or SCD, in patients with ICM.
Methods
In the multicenter DERIVATE–ICM registry (N=861), we developed a CMR-based ML risk score (ICM-RiskML) to predict MAACE. Patients were stratified by center into training (N=639) and testing (N=222) cohorts. Feature selection was performed using both Random Survival Forest (RSF) and LASSO-penalized Cox regression. Among 132 baseline variables, 25 predictors were retained for feature selection (Figure 1). Six variables: LVEF, New York Heart Association (NYHA) class, indexed left ventricular end-diastolic volume (LVEDVi), left atrial area, dense late gadolinium enhancement (LGE), and non-dense LGE—were consistently selected by both methods. Final prediction models were trained using RSF and Cox on the selected predictors. Internal model development and hyperparameter tuning were performed using nested cross-validation within the training cohort. Model performance was assessed using Uno’s concordance index, accuracy, specificity, sensitivity, and Kaplan-Meier risk stratification in the independent testing cohort.
Results
Over a median follow-up of 2.7 years, 88 patients (10%) experienced MAACE (58 in the training set, 30 in the testing set). The overall cohort was predominantly male (86%), with a mean age of 65 ± 11 years. Most patients were in NYHA class II–III (61%), and the mean LVEF was 28±10%, with a mean LVEDVi of 124±43 mL/m². In the testing cohort, RSF outperformed Cox regression (C-index 0.69 [0.68–0.71] vs 0.61 [0.59–0.62], p<0.001). Compared with guideline-based ICD allocation, ICM-RiskML showed higher specificity (0.78 vs 0.38) and higher overall accuracy (0.67 vs 0.56, p<0.001). Kaplan–Meier analyses demonstrated clear risk stratification across ICM-RiskML categories in the overall population and in patients with LVEF ≤35% (Figure 2).
Conclusion
A novel CMR-based ML risk score improved arrhythmic risk stratification compared with guideline-based ICD allocation. The ability of ICM-RiskML to discriminate risk among patients with LVEF≤35% supports a higher specificity than LVEF-based criteria alone. Integrating CMR-derived predictors into ML models may help refine prophylactic ICD decision-making and warrants external validation.Fig 1.Variables selectionSurvival curves