Topological Data Analysis Approach to Extract the Persistent Homology Features of Ballistocardiogram Signal in Unobstructive Atrial Fibrillation Detection

• Published in: IEEE sensors journal Vol. 22; no. 7; pp. 6920 - 6930
• Main Authors: Jiang, Fangfang, Xu, Bowen, Zhu, Ziyu, Zhang, Biyong
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Arrhythmia; atrial fibrillation; Ballistocardiogram signal; Ballistocardiograms; Bar codes; Cardiac arrhythmia; Classification; Data analysis; Diagnosis; Electrocardiography; Feature extraction; Fibrillation; Homology; Machine learning; Nonlinear dynamics; persistent homology features; phase space reconstruction; Rhythm; Sensors; Time series analysis; Time-frequency analysis; topological data analysis; Topology; Trajectory; unobstructive detection
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2022.3153647

Atrial fibrillation (AF) is the most common arrhythmia, which requires long-term diagnosis and treatment in a daily routine. Recently, ballistocardiogram (BCG), an unobstructive cardiac function monitoring method, is widely studied for AF detection. Usually, the AF classification performance depends on the characterization property of extracted features. Therefore, we propose a novel nonlinear persistent homology feature, combining theory from nonlinear dynamics and persistent homology based topological methods, which is aimed at providing supplementary to existing features and further improving AF diagnosis performance. In this research, one-dimension BCG series were reconstructed to high-dimension phase space point clouds firstly, which contain more abundant rhythm information. Then topological distributions of the point clouds were represented as persistent homology barcodes using topological data analysis. Finally, the statistics of the barcodes were considered as 9 persistent homology features to quantify the barcodes. To validate the proposed AF detection method, we collected 4000 AF and non-AF segments of BCG from 73 subjects, and 6 machine learning classifiers were performed. By combining the 9 persistent homology features with 17 previously proposed features, our features brought a 6.17% increment in accuracy compared with the 17 features solely (<inline-formula> <tex-math notation="LaTeX">{p} < 0.001 </tex-math></inline-formula>), reaching 94.50%. Using feature selection technique, 12 effective features were reserved and achieves 93.50% classification accuracy. It follows that the proposed features contribute to improving AF detection performance for a larger amount of BCG data, which contains versatile pathological information and individual differences.