Multivariate ensemble classification for the prediction of symptoms in patients with Brugada syndrome.

Identification of asymptomatic patients at higher risk for suffering cardiac events remains controversial and challenging in Brugada syndrome (BS). In this work, we proposed an ECG-based classifier to predict BS-related symptoms, by merging the most predictive electrophysiological features derived from the ventricular depolarization and repolarization periods, along with autonomic-related markers. The initial feature space included local and dynamic ECG markers, assessed during a physical exercise test performed in 110 BS patients (25 symptomatic). Morphological, temporal and spatial properties quantifying the ECG dynamic response to exercise and recovery were considered. Our model was obtained by proposing a two-stage feature selection process that combined a resampled-based regularization approach with a wrapper model assessment for balancing, simplicity and performance. For the classification step, an ensemble was constructed by several logistic regression base classifiers, whose outputs were fused using a performance-based weighted average. The most relevant predictors corresponded to the repolarization interval, followed by two autonomic markers and two other makers of depolarization dynamics. Our classifier allowed for the identification of novel symptom-related markers from autonomic and dynamic ECG responses during exercise testing, suggesting the need for multifactorial risk stratification approaches in order to predict future cardiac events in asymptomatic BS patients. Graphical abstract Pipeline for feature selection and predictive modeling of symptoms in Brugada syndrome.