Prediction of ventricular arrhythmias in non-ischemic cardiomyopathy using a 12-lead ECG-based machine learning model
M Garcia-Montero, M Bahani, E Deschamps, X Y Xie, S Djerroud, M Tremblay Gravel, L P David, S Hermann Honfo, R Tadros, R Avram, J Cadrin-TourignyAbstract
Background
Arrhythmic risk stratification in non-ischemic cardiomyopathy (NICM) remains challenging. Current predictors, such as left ventricular ejection fraction (LVEF) <35%, late gadolinium enhancement (LGE) and genetic markers have limitations in accurately identifying patients at risk for sustained ventricular arrhythmias (SVA). Our aim was to develop an ECG-based machine learning (ECG-ML) model to predict SVA in patients with NICM and to compare its predictive ability with established risk factors.
Methods
In this retrospective, single-center study, we included patients with NICM and no history of SVA, who had undergone baseline digitalized 12-lead ECG, genetic testing, and cardiac magnetic resonance. The primary outcome was SVA, defined as sudden cardiac death, sustained ventricular tachycardia or appropriate ICD intervention. A neural network–based model was developed using ECGs from 80% patients of the cohort (training cohort). Two model architectures were evaluated: an ECG-only model (ECG-ML) and a combined ECG-and-clinical model integrating ECG features with the patient's age at diagnosis, LVEF, the presence and type of LGE, and the presence of a pathogenic variant in a NICM-related or high arrhythmic-risk gene. Model performance was assessed in a separate validation cohort comprising the remaining 20% of patients. Results were compared with a clinical reference model incorporating the aforementioned clinical risk factors using Least Absolute Shrinkage and Selection Operator (LASSO) multivariate logistic regression.
Results
A total of 314 patients were included (68.5% male). The median age at diagnosis was 50.4 years (IQR 40.1–58.7), and the median LVEF was 28% (IQR 19–40%). LGE was present in 63.7% of patients, with 36.2% meeting high-risk criteria. Genetic testing was positive in 23% of patients, and 3.18% harbored a causal variant in high-risk genes. During a median follow-up of 3.91 years (IQR 2.27–4.58), 7.96% of patients developed SVA. The best-performing model was the ECG-ML model, which predicted SVA (p < 0.001) with a sensitivity of 75%, specificity of 85%, and an AUC of 0.86 (95% CI 0.57–1.00) in the validation cohort. No clinical covariate improved the ECG-ML model's performance. The clinical reference model, including LGE presence, high-risk LGE patterns, and a positive genotype, showed limited discriminative ability for predicting incident SVA (AUC 0.53; 95% CI 0.24–0.82). Neither LVEF nor the presence of a high-risk genotype were independent predictors of SVA in our cohort.
Conclusion
The ECG-based machine learning (ECG-ML) model demonstrated the potential to predict incident SVA in patients with NICM, with performance potentially surpassing that of conventional clinical variables. However, these findings are derived from a limited dataset and warrant confirmation in larger cohorts for external validation.Baseline characteristics of the cohortROC curves for ECG-ML and combined model