Agent-based beat-by-beat compression of 12-lead electrocardiogram signal using adaptive Fourier decomposition. (May 2022)

• Record Type: Journal Article
• Authors: Banerjee, Soumyendu; Singh, Girish Kumar
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: Multilead electrocardiogram -- Agent-based compression -- Möbius transform -- Adaptive Fourier decomposition -- Optimal decomposition level; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103628
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Agent based beat-wise 12-lead ECG compression. Principal component analysis-based dimensionality reduction in nonlinear domain. Implementation of Möbius Transform-based Adaptive Fourier Decomposition. Multilayer perceptron neural network based optimal decomposition in AFD. Average CR and PRD obtained of 48.21 and 3.88 respectively, tested upon 546 PTBDB records. Abstract: Objective: The multilead electrocardiogram (MECG) signal provides much more detailed information related to the cardiac activity of the human heart than single channel signal. Long-term recording of MECG data needs a huge amount of storage space, thus data compression becomes truly necessary. Methods: In this work, a beat-wise MECG data compression is proposed that is based on adaptive Fourier decomposition (AFD). To reduce dimensionality, an ECG beat was treated as a multiagent, upon which principal component (PC) analysis was used in non-linear space. A new Möbius transform was introduced along with AFD, to convert the dominant PCs in complex domain using Nevanlinna factorization. An offline trained multilayer perceptron neural network was also employed to provide optimal decomposition levels in AFD to limit the PRD within 3%. Result: The entire work was tested on 546 ptbdb MECG records available in Physionet which yielded an average compression ratio and a percent root mean squared difference (PRD) of 48.21 and 3.88, respectively. Regardless of annotation type, PRD variance within each lead was found to be nearly uniform. Conclusion: Within a single beat, the agent-based compression established a hybrid compression technique, almost completely preserving critical clinical aspects. Significance: Beat-wise compression, therefore requires less buffer memory, allowing it to be used for real-time data compression. The low reconstruction error enhanced the acceptability of proposed work in medical applications.