Poincaré Image Analysis of Short-Term Electrocardiogram for Detecting Atrial Fibrillation.

Abstract

Atrial fibrillation (AF) is a heart rhythm disorder and is associated with the risk of stroke and heart failure. Early detection of AF is crucial but challenging due to its asymptomatic nature and similarity to other ectopic beats, such as premature atrial contractions (PACs) and premature ventricular contractions (PVCs). This article presents a novel Poincaré image-domain feature-based automated AF screening model in the presence of PACs/PVCs using 10-second single-lead electrocardiogram (ECG) signals. The model proposes a hybrid approach that integrates a radial basis function-based support vector machine classifier, optimized via grid search, with a rule-based decision criterion. A set of 84 Poincaré image features is extracted and reduced to a set of four features through the minimum redundancy maximum relevance selection approach and then fed into the classifier. Additionally, rules based on P-wave information and dRR distribution patterns of ECG signal are incorporated to enable a more distinct separation of PACs/PVCs from AF. The model was validated using eight datasets comprising recordings from 25,776 subjects. Both 5-fold cross-validation and leave-one-dataset-out validation were performed using 2,06,367 segments: 1,12,591 normal, 9,485 PACs/PVCs, and 84,291 AF segments. The accuracy ranges were 96.35% to 99.40% and 96.48% to 99.33% for 5-fold cross-validation and leave-one-dataset-out validation, respectively, with balanced sensitivity and specificity across all datasets. The model's superior performance across diverse data demonstrates its robustness and suitability for real-world application, supporting its potential in computerized assessment of short-term ECGs to detect AF.