Pairwise Feature Interactions to Predict Arrhythmic Risk of Brugada Syndrome

Sharen Lee1, Jiandong Zhou2, Konstantinos Letsas3, Christien Li4, Tong Liu5, Sven Zumhagen6, Eric Schulze-Bahr6, Gary Tse5, Qingpeng Zhang2
1Laboratory of Cardiovascular Physiology, Li Ka Shing Institute of Health Sciences, Hong Kong, SAR, P.R. China, 2City University of Hong Kong, 3Evangelismos General Hospital of Athens, 4Newcastle University, 5Second Hospital of Tianjin Medical University, 6Hospital M√ľnster


Abstract

Introduction: Electrocardiographic (ECG) indices of depolarization and repolarization were used for risk stratification in Brugada syndrome (BrS). Nonlinear interaction patterns between risk factors were ignored for reliable risk prediction.

Methods: We adapted a generalized additive model with pair-wise interactions (GA2M) to predict BrS with spontaneous ventricular tachycardia/fibrillation (VT/VF) as outcomes based on specific ECG markers. The GA2M considers flexible nonlinearities among risk factors, and still retains the simple additive structure of linear models for intuitive explanations. A total of 191 adult patients with BrS from three centres (Germany, Greece and Hong Kong) were included for analysis. Depolarization and repolarization ECG markers were measured from the right precordial leads (V1 to V3).

Results: The proposed GA2M-based risk prediction model successfully (a) identified a set of risk factors and their pair-wise interactions (e.g. between dispersion of conduction (QRS dispersion) and dispersion of repolarization/total repolarization (Tpeak-Tend x mean QT)), (b) revealed a set of pair-wise interaction patterns that have been ignored by previous studies due to the limitation of linear models, and (c) outperformed the baseline logistic model based on the same set of ECG measurements.

Conclusions: The findings implicate pairwise interactions between dispersion of conduction and repolarization abnormalities in the arrhythmic substrate. These patterns can be extracted by GA2M and their inclusion improved predictive performance and enabled more effective risk stratification in BrS.