Explanation of HRV Features for Detecting Atrial Fibrillation

Abstract

Atrial fibrillation is a heart condition that is known to be affecting approximately 33 million people in the world and cause a wide range of complications, that may at times be fatal if not properly treated. Many studies try to address this condition by either training highly accurate artificial intelligence detector models or by explaining the physiological significance of certain features used to train these detector models. However, the problem with these kind of studies is that it leaves its users to either blindly trust the calls of a black box model, or merely inform themselves of various features. In this study, we try to address this problem by revealing which features are most frequently chosen by artificial intelligence models to detect atrial fibrillation and also try to explain why they might be so significant. To achieve this, heart rate variability features are extracted from the ‘PhysioNet/Computing in Cardiology (CinC) Challenge 2017’ dataset and are used to train machine learning models like Random Forest and XGBoost, with Recursive Feature Elimination. We find that many machine learning algorithms find the ‘hr_std’, ‘sd_ratio’, ‘nni_mean’, and ‘pnn20’ features to be most important in detecting atrial fibrillation. We achieve a mean F$_1$-score of 98.13% and 95.18% for Random Forest and Support Vector Machine with a Linear Kernel, respectively, using K-Fold Cross-Validation with five-folds.