Electrocardiogram arrhythmia detection with novel signal processing and persistent homology-derived predictors

Abstract

Many approaches to computer-aided electrocardiogram (ECG) arrhythmia detection have been performed, several of which combine persistent homology and machine learning. We present a novel ECG signal processing pipeline and method of constructing predictor variables for use in statistical models. Specifically, we introduce an isoelectric baseline to yield non-trivial topological features corresponding to the P, Q, S, and T-waves (if they exist) and utilize the N-most persistent 1-dimensional homological features and their corresponding area-minimal cycle representatives to construct predictor variables derived from the persistent homology of the ECG signal for some choice of N. The binary classification of (1) Atrial Fibrillation vs. Non-Atrial Fibrillation, (2) Arrhythmia vs. Normal Sinus Rhythm, and (3) Arrhythmias with Morphological Changes vs. Sinus Rhythm with Bradycardia and Tachycardia Treated as Non-Arrhythmia was performed using Logistic Regression, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Naive Bayes, Random Forest, Gradient Boosted Decision Tree, K-Nearest Neighbors, and Support Vector Machine with a linear, radial, and polynomial kernel Models with stratified 5-fold cross validation. The Gradient Boosted Decision Tree Model attained the best results with a mean F1-score and mean Accuracy of ( 0.967 , 0.946 ), ( 0.839 , 0.946 ), and ( 0.943 , 0.921 ) across the five folds for binary classifications of (1), (2), and (3), respectively.