Interpretable machine learning model using Extreme Gradient Boosting with Nonnegative Matrix Factorization improves the accuracy of arrhythmic risk prediction in Brugada Syndrome