Artikel
Quantifying Baseline Noise in 12-Lead ECG
Suche in Medline nach
Autoren
Veröffentlicht: | 15. September 2023 |
---|
Gliederung
Text
Introduction: Electrocardiograms (ECGs) suffer from artifacts of diverse sources that might reduce signal-to-noise ratio (SNR) to a level that manual interpretation is time-consuming or prone to misdiagnosis. In clinical routine, staff can recognize these artifacts during acquisition and decide whether signal quality is adequate for clinical interpretation or new measurements are necessary. However, with the rise of long-time ECG wearables this manual analysis is not feasible anymore. Thereby, computer algorithms for quality assessment are required, but often tend towards false assessments when failing to differentiate between signal and artifact.
In this work, we focus on baseline drift, which is a challenging artifact as it might result from various sources, e.g. unwanted physiological effects such as respiration but also inadequate electrode contact. Furthermore, it has a time-varying characteristic and might change its intensity over time. Therefore, its automatic detection is a challenge and many publicly-available datasets only contain qualitative manual annotations of small subsets [1].
We apply a baseline detector, which has proven valuable in chromatography, to a large-scale 12-lead ECG database. By proposing a SNR measure, we quantify the amount of baseline drift within each signal.
Methods: We developed a Python algorithm to analyze 10s 12-lead ECGs from the public database PTB-XL (sampling rate: 500 Hz) [2]. The algorithm input is a single ECG and each lead is processed independently: The baseline is estimated using BEADS methods (https://pypi.org/project/pybeads/ v.1.0.1), based on an optimization problem where the baseline is modeled as a low-pass signal and peaks are assumed as being sparse [3].
Subsequently, SNR is computed by assuming the baseline signal as “noise”. By computing the variance of both noise signal and original signal with subtracted noise, computing their ratio, taking its logarithm to the base 10 and multiplying with 10 afterwards, we arrive at SNR in decibels.
Results: We applied the algorithm to 1598 ECGs of the PTB-XL, which have a label denoting that baseline drift is present. As we process the ECGs lead-wise, this results in 19176 SNR values. SNR has a range of [-29 dB, 26 dB] with an average of 3.5 dB (std: 7.1 dB).
Subsequently, we conducted visual analysis of manually-selected results. For negative SNR values, ECGs show pronounced baseline drifts, deviating severely from the isoelectric line. We observed individual cases where the SNR did not represent the baseline drift correctly. This occurred in ECGs showing QRS complexes with very low amplitudes or PQ- or ST-segment elevation.
Discussion: Results suggest that the manual annotations of PTB-XL cover a wide range of SNR values, demonstrating the need for a more quantitative measure. Visual analysis suggests that the BEADS algorithm works well, which is in line with other works [4]. However, abnormalities affecting the isoelectric line led to biased results in certain cases, warranting future research.
Conclusion: The proposed SNR metric results in a meaningful quality indicator for baseline drift and is a promising measure for representing baseline drifts quantitatively. In future work, we will use this information to evaluate the influence of baseline noise on ECG classification using Deep Neural Networks [5].
The authors declare that they have no competing interests.
The authors declare that an ethics committee vote is not required.
References
- 1.
- Bender T, Gemke P, Idrobo-Avila E, Dathe H, Krefting D, Spicher N. Benchmarking the Impact of Noise on Deep Learning-based Classification of Atrial Fibrillation in 12-Lead ECG [Preprint]. arXiv. 2023. DOI: 10.48550/arXiv.2303.13915
- 2.
- Wagner P, Strodthoff N, Bousseljot RD, Samek W, Schaeffter T. PTB-XL, a large publicly available electrocardiography dataset. PhysioNet; 2022. DOI: 10.13026/kfzx-aw45
- 3.
- Ning X, Selesnick IW, Duval L. Chromatogram baseline estimation and denoising using sparsity (BEADS). Chemometrics and Intelligent Laboratory Systems. 2014;139:156–67. DOI: 10.1016/j.chemolab.2014.09.014
- 4.
- Wang X, Zhou Y, Shu M, Wang Y, Dong A. ECG Baseline Wander Correction and Denoising Based on Sparsity. IEEE Access. 2019;7:31573–85. DOI: 10.1109/ACCESS.2019.2902616
- 5.
- Bender T, Beinecke JM, Krefting D, Muller C, Dathe H, Seidler T, Spicher N, Hauschild AC. Analysis of a Deep Learning Model for 12-Lead ECG Classification Reveals Learned Features Similar to Diagnostic Criteria. IEEE J Biomed Health Inform. 2023 May 1. DOI: 10.1109/JBHI.2023.3271858