Learning-Based Approach to Predict Fatal Events in Brugada Syndrome

Brugada Syndrome (BrS) is an arrhythmic disorder which increases the probability of developing arrhythmic events, even life-threatening ones, in young and otherwise healthy individuals. It accounts for 5–20% of sudden deaths in people with no structural cardiac abnormality. The first clinical manifestation of this syndrome is, usually, a cardiac arrest. A correct evaluation of the risk of developing an arrhythmic event could prevent premature deaths and unnecessary procedures. This paper focuses on the idea that analysis based on machine learning can extract some piece of information not visible to the human eye and, thus, forecast if a sudden death will occur or not. The study population comprises 209 electrocardiograms (ECGs) from the Piedmont Brugada register, 41 of subjects who had an event, while the remaining 168 are used as controls; therefore, it is a binary classification problem. Cardiologists manually measured 24 features per ECG. Then, a multi-layer perceptron (MLP), a boosted decision tree (BDT) model, a decision tree, a Support Vector Machine (SVM), and a Naïve Bayes (NB) classifier were employed to classify the ECGs. All models show a high negative predictive value: a patient whose predicted class is negative is likely to remain asymptomatic. Since the positive predictive values of the MLP and NB are not sufficiently high, the opposite cannot be stated. Finally, F1-score shows BDT outperforms (0.67) all the other models.