Selecting an Optimal Training Dataset for Machine Learning based Atrial Fibrillation Detection

Abstract

The application of Machine Learning, in recent times, has ex- celled with positive outcome in many fields, including the medical field, such as handling cardiovascular problems. In this paper, we aim at de- veloping a machine learning algorithm for detecting Atrial Fibrillation, as one of the most common and mortal types of heart rhythm prob- lems - arrhythmias. Especially we address the research question of which dataset to be used in the learning process to reveal optimal results. The experiments are conducted using the following algorithms: Support Vec- tor Machines, Decision Trees and Random Forest training validating and testing on specific selection of the three most popular publicly available electrocardiogram databases that contain episodes of Atrial Fibrillation. The research concluded that the best results are obtained by the Random Forest algorithm trained on LTAFDB selected by the 80-10-10 rule for training, validation and testing.