Electrocardiogram Reconstruction Based on Compressed Sensing

Zhimin Zhang, Xinwen Liu, Shoushui Wei, Hongping Gan, Feifei Liu, Yuwen Li, Chengyu Liu, Feng LiuIEEE Access 2019 · DOI

Abstract

The proliferation of wearable cardiac monitors necessitates efficient data acquisition and transmission to preserve battery life. Traditional sampling methods following the Shannon-Nyquist theorem generate significant data redundancy, leading to high power consumption. This paper proposes a Compressed Sensing (CS) framework that leverages the quasi-periodic nature of electrocardiogram (ECG) signals. By mapping 1D signals into a 2D space and employing block-sparse recovery algorithms, the researchers achieved clinical-grade reconstruction using only 30% of the data required by traditional methods.

Introduction

The gold standard for heart monitoring, the 12-lead ECG, is traditionally bulky and power-intensive. As healthcare shifts toward continuous, mobile monitoring via wearables, battery life and data transmission bandwidth have become critical bottlenecks. Standard "sample-then-compress" approaches, based on the Shannon-Nyquist theorem, sample at twice the signal's bandwidth, resulting in massive amounts of redundant data. This paper addresses these challenges by applying Compressed Sensing (CS), a "compress-while-sampling" paradigm that captures essential information directly, significantly reducing the energy burden on acquisition and wireless transmission.

Methodology

The researchers introduced two primary schemes for ECG processing:

1. 2D Under-sampling Scheme

To exploit the quasi-periodic nature of ECG signals, the authors developed the Cut and Align (CAB) method. This technique segments 1D signals based on R-peaks and aligns them into a 2D array. This mapping captures both intra-beat (within a single cycle) and inter-beat (between adjacent cycles) correlations, enhancing signal sparsity in the transform domain.

Transform domain sparsity showing the concentration of signal energy.

Furthermore, a "Region of Interest" (RoI) sampling approach was implemented. Instead of uniform random sampling, the system samples more densely around the QRS complex—the most clinically significant portion of the heartbeat—and more sparsely elsewhere. Reconstruction is then achieved using a conjugate gradient optimization model to solve the $l_1$-norm minimization problem.

2. 1D Compression Scheme

In this alternative, the signal is compressed using a Gaussian random matrix. The study compared four recovery algorithms: Matching Pursuit (MP) variants (OMP and CoSaMP) and Block Sparse Bayesian Learning (BSBL) variants (BO_BSBL and EM_BSBL).

Key Insights: Structure is Sparsity

The core innovation lies in the realization that "structure is sparsity." By transforming the 1D time series into an image-like 2D structure through heartbeat alignment, the researchers created a representation where the background is highly redundant and the signal features are concentrated.

2D mapping of aligned ECG heartbeats, highlighting the structural redundancy.

The results also emphasize the superiority of BSBL for biological signals. Because ECG features like the QRS complex appear in clusters or "blocks," BSBL algorithms—designed to exploit intra-block correlation—outperform standard MP methods, especially at high compression ratios.

Results and Discussion

Using data from the MIT-BIH Normal Sinus Rhythm Database (NSRDB), the study demonstrated that the 2D scheme could reconstruct signals with only 30% of the original data while maintaining a Percentage Root-mean-square Difference (PRD) below 9%, the threshold for clinical quality.

In the 1D scheme, the BO_BSBL algorithm proved most effective, maintaining accuracy even at low sampling frequencies where MP methods failed. Crucially, reconstruction for a 10-second ECG segment was consistently completed in under 10 seconds, meeting the real-time requirements for clinical monitoring.

Conclusion

This CS-based framework offers a viable path for long-term, high-fidelity cardiac monitoring on power-constrained wearable devices. While the computational complexity of BSBL and the reliance on accurate R-peak detection for 2D mapping present ongoing challenges, the significant reduction in data acquisition requirements marks a major step toward "anywhere, anytime" heart health monitoring. Future work will focus on optimizing algorithm efficiency and testing against diverse cardiac pathologies.

Summary updated

BibTeX
@article{zhang2019electrocardiogram,
  author = {Zhimin Zhang and Xinwen Liu and Shoushui Wei and Hongping Gan and Feifei Liu and Yuwen Li and Chengyu Liu and Feng Liu},
  title = {{Electrocardiogram Reconstruction Based on Compressed Sensing}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2905000},
}
Figure