Machine Learning-Based Clustering Using a 12-Lead Electrocardiogram in Patients With a Implantable Cardioverter Defibrillator to Identify Future Ventricular Arrhythmia.

Abstract

Background: Implantable cardioverter defibrillators (ICDs) reduce mortality associated with ventricular arrhythmia in high-risk patients with cardiovascular disease. Machine learning (ML) approaches are promising tools in arrhythmia research; however, their application in predicting ventricular arrhythmias in patients with ICDs remains unexplored. We aimed to predict and stratify ventricular arrhythmias requiring ICD therapy using 12-lead electrocardiograms (ECGs) in patients with an ICD.

Methods and results: This retrospective analysis included 200 adult patients who underwent ICD implantation at a single center. Patient demographics, clinical features, and 12-lead ECG data were collected. Unsupervised learning techniques, including K-means and hierarchical clustering, were used to stratify patients based on 12-lead ECG features. Dimensionality reduction methods were also used to optimize clustering accuracy. The silhouette coefficient was used to determine the optimal method and number of clusters. Of the 200 patients, 59 (29.5%) received appropriate therapy. The mean age of patients was 62.3 years, and 81.0% were male. The mean follow-up period was 2,953 days, with no significant intergroup differences. Hierarchical clustering into 3 clusters proved to be the most accurate (silhouette coefficient=0.585). Kaplan-Meier curves for these 3 clusters revealed significant differences (P=0.026).

Conclusions: We highlight the potential of ML-based clustering using 12-lead ECGs to help in the risk stratification of ventricular arrhythmia. Future research in a larger multicenter setting may provide further insights and refine ICD indications.