Analyzing the performance of biomedical time-series segmentation with electrophysiology data

Figures

• 

  Fig. 1. The Bayesian optimization workflow. Function f(x; P) represents any of the analysed methods, each involving its unique intrinsic set of hyperparameters denoted as P. P is iteratively optimised to maximise quality metric Q representing the harmonic mean of precision and recall (Dice score). Q is constrained to a range between 0 and 1.

• 

  Fig. 2. Comparison of B-EGM segmentation performance for sinus rhythm. The figure illustrates the detection and delineation of atrial complexes in the presence of normal atrial rhythm and ventricular far-field potentials. The ground truth segments are marked with gray rectangles. The color-coded rectangles represent the segmentation results from the rule-based (green), support vector machine (red), and deep neural network methods based on U-Net (blue), R-CNN (violet), and DENS-ECG (cyan), respectively. The SVM and DNN methods demonstrated superior inter-beat performance in accurately identifying complex boundaries, which is essential for reproducibility of diagnostic protocols. Signal clipping marked by dashed lines was applied to prevent a high voltage signals from overflowing outside the plotted area.

• 

  Fig. 3. Comparison of B-EGM segmentation performance for sinus rhythm with false positive detection. The figure illustrates the detection and delineation of low-amplitude atrial complexes recorded by catheter with decreased tissue-electrode contact. The DNN methods demonstrated superior inter-beat performance, while SVM falsely misclassified ventricular far-field potential as atrial complex. The ground truth segments are marked with gray rectangles. The color-coded rectangles represent the segmentation results from the rule-based (green), support vector machine (red), and deep neural network methods based on U-Net (blue), R-CNN (violet), and DENS-ECG (cyan), respectively. Signal clipping marked by dashed lines was applied to prevent a high voltage signals from overflowing outside the plotted area.

• 

  Fig. 4. Comparison of B-EGM segmentation performance for paced rhythm. The figure illustrates the detection and delineation of high-amplitude ventricular far-field potentials and atrial complexes fused with pacing artifacts due to the artificially paced rhythm. The SVM and R-CNN methods overestimated the beginning of the complex due to the presence of pacing artifacts. The rule-based method also mistakes prominent far-field activity as an atrial complexes. The ground truth segments are marked with gray rectangles. The color-coded rectangles represent the segmentation results from the rule-based (green), support vector machine (red), and deep neural network methods based on U-Net (blue), R-CNN (violet), and DENS-ECG (cyan), respectively. Signal clipping marked by dashed lines was applied to prevent a high voltage signals from overflowing outside the plotted area.

• 

  Fig. 5. Comparison of B-EGM segmentation performance for reentry tachycardia. The figure illustrates the detection and delineation of superimposed atrial complexes and ventricular far-field potentials (first three complexes) during an ongoing tachycardia. The ground truth segments are marked with gray rectangles. The color-coded rectangles represent the segmentation results from the rule-based (green), support vector machine (red), and deep neural network methods based on U-Net (blue), R-CNN (violet), and DENS-ECG (cyan), respectively. Each method demonstrated capability of identifying superimposed atrial complexes, which is essential for interval measurement and arrhythmia mapping during electrophysiology procedure. Signal clipping marked by dashed lines was applied to prevent a high voltage signals from overflowing outside the plotted area.

• 

  Fig. 6. Detection error distribution. Violin plots showing the distribution of the detection error for the onset (left part) and offset (right part) detection methods. As observed, the U-Net and DenseNet methods exhibit the smallest variability, indicating it is the most stable method for both onset and offset detection.

• 

  Fig. 7. Models’ performance for GWN noise. The trend of the monitored metrics (Dice score, recall and precision) as a function of increasing GWN level for each method. The rule-based method is very sensitive to increasing noise levels and starts to gradually fail even with low noise addition.

• 

  Fig. 8. Models’ performance for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${PLI}_{50}$$\end{document} noise. The trend of the monitored metrics (Dice score, recall and precision) as a function of increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PLI_{50}$$\end{document} level for each method. In this comparison, the SVM method performs worst, as it becomes unreliable at higher noise intensities and fails to correctly detect atrial activity in the signals.

• 

  Fig. 9. Models’ detection error for GWN noise. The trend of the ME and RMSE as a function of increasing GWN level for each method. Similar to Figure 7, it can be seen that the rule-based method exhibits considerable variability in the resulting delineation with increasing noise intensities.

• 

  Fig. 10. Models’ detection error for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${PLI}_{50}$$\end{document} noise. The trend of the ME and RMSE as a function of increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$PLI_{50}$$\end{document} level for each method. Compared to Figure 9, which shows considerable variability in the outputs of the rule-based method, there is not much change in the SVM model in this case. Thus, in the context of Figure 8, it can be said that the SVM method only has a detection problem in the noisy data, not in the delineation of the atrial activity found.