Supervised Classification of Brugada Syndrome Patients by ECG-Derived Markers

Abstract

Brugada syndrome (BrS) has been associated with risk of ventricular fibrillation and sudden cardiac death (SCD). Its risk stratification remains challenging as the only accepted factor is the presence of resuscitated cardiac arrest or arrhythmogenic syncope and the majority of patients are diagnosed in the asymptomatic phase. Moreover, the only treatment available to prevent SCD is the implantation of a cardiac defibrillator, which can lead to adverse events such as inappropriate shocks. In this study, we present Machine Learning (ML)/supervised classification tools for BrS risk stratification based on the automatic analysis of long-term high-resolution electrocar-diographic information. For this purpose, 12-lead ECG 24h Holter and clinical variables from 64 Brugada subjects were used. ECG signals were preprocessed with a signal-averaging algorithm to reduce noise and obtain individual ECG beats for delineation, resulting in 11 ECG biomarkers. Subsequently, 4 different ML/supervised algorithms based on Decision Tree, XGBoost, K-Nearest Neighbors and support vector machine algorithms were tested. AUC results were around 90%, however sensitivity results were around 50%. The results do not efficiently predict BrS symptomatic patients at risk of SCD, which is mainly caused by the reduced number of symptomatic patients. Further studies with additional subjects and variables may improve this prognosis.