Learning-based approach to predict fatal events in Brugada Syndrome

Learning-based approach to risk stratification in Brugada Syndrome threatening arrhythmic disorder characterized by an increased probability to develop arrhythmic events such as ventricular tachycardia and fibrillation in young and otherwise healthy individuals. It is responsible for 5 to 40% of sudden deaths in patients without structural heart abnormalities. The global prevalence of BrS ranges from 5 to 20 cases in every 10000 individuals worldwide, with particular incidence in Asian population. Cardiac arrest is often the first clinical manifestation of the disease in previously asymptomatic patients. Therefore, the need to assess a robust method to define the risk of developing an arrhythmic event associated with a patient is particularly evident. A correct evaluation of the arrhythmogenic potential could lead to correct therapeutical decisions and, therefore, prevent premature deaths and unnecessary procedures. The primary prevention tool in BrS is the implantable cardiac defibrillator (ICD). However, whether to use this tool is controversial because the risk of sudden cardiac death (SCD) among young patients is relatively low (0.5–1.5% per year) while the rate of complications related to ICD insertion is rather high, especially at a young age. So, the main challenge for physicians is to identify patients at risk of developing arrhythmia and therefore require specific treatment. Many risk factors are associated with the increase in the probability of develop- ing an arrhythmic event for previously asymptomatic patients, such as age, presence of genetic mutations, family history, and various ECG intervals and measurements. The published reviews regarding these risk factors show conflicting data. At present, there is no commonly assessed risk stratification method. This thesis project focuses on the idea that the ECG, the election diagnostic tool for BrS, may contain information about the risk of developing a life-threatening arrhythmic event, information that is not visible to the human eye and knowledge. In particular, the underlying hypothesis is that machine learning-enhanced analysis can retrieve this information and correctly predict whether a patient will develop an event or not. The population used to test machine learning models was selected from Regione Piemonte’s Brugada register and comprises measurements from 209 ECG traces, 41 of which belonged to patients who developed an event and 168 controls. A total of 24 features were measured manually by cardiologists. The learning problem is a binary classification problem, and the models developed are a boosted decision tree model and a multi-layer perceptron network. The performances of these models were confronted with a simple decision tree model, a naïve Bayes classifier and a support vector machine model. The models all show a high negative predictive value, which looks like a promising result: the result of the analysis tells the cardiologist that a patient whose predicted class is negative (the model predicts he/she will not have an event) is likely to remain asymptomatic, therefore not needing the implantation of an ICD. The positive predictive value is relatively low, so the vice versa cannot be inferred.