Robust T-End Detection via T-End Signal Quality Index and Optimal Shrinkage

Figures

• 

  Figure 1. The comparison of bSQI and tSQI. The black line is the ECG signal from the Ultra-Long-Term ECG recordings Database with intermediate quality. The intersection of x-axis and blue (magenta) vertical lines are the time stamps of R peak locations estimated by Elgendi’s algorithm [34] (jqrs [30]). The intersection of x-axis and red (green) vertical lines are the time stamps of T-end locations estimated by Zhang’s algorithm (Martinez’s algorithm). The magenta vertical lines are shifted right by 10 milliseconds to avoid the overlap with blue lines. The value of bSQI is 0.99, and the value of tSQI is 0.66, which are calculated by matching points from two different algorithms as the same point if the time difference between them is smaller than 50 ms.

• 

  Figure 2. Comparison of different T-end detection algorithms with or without optimal shrinkage when the signal is contaminated by the additive Gaussian noise. In each subplot, the black line is the ECG signal from the first channel of a subject in the QT database with additive noise, the light blue line is the denoised ECG signal by the optimal shrinkage, and the magenta line is the original ECG signal without adding any noise. The red (green) vertical lines indicate the estimated T-end by the chosen detection algorithm on noisy signals without (with) OS applied. The deep blue vertical line indicates the annotated T-end by Cardio 1. The plots in the first (second) row are the results when the SNR is 10 dB (5 dB). The chosen T-end algorithm in the first, second and third column is Zhang’s, Carlos’s, and Martinez’s algorithms respectively. The case index of each plot is in the title.

• 

  Figure 3. Comparison of different T-end detection algorithms with or without optimal shrinkage when the signal is contaminated by the additive ARMA (1,1) noise. In each subplot, the black line is the ECG signal from the first channel of a subject in the QT database with additive noise, the light blue line is the denoised ECG signal by the optimal shrinkage, and the magenta line is the original ECG signal without adding any noise. The red (green) vertical lines indicate the estimated T-end by the chosen detection algorithm on noisy signals without (with) OS applied. The deep blue vertical line indicates the annotated T-end by Cardio 1. The plots in the first (second) row are the results when the SNR is 10 dB (5 dB). The chosen T-end algorithm in the first, second and third column is Zhang’s, Carlos’s, and Martinez’s algorithms respectively. The case index of each plot is in the title.

• 

  Figure 4. Scatterplot of T-end prediction error before and after applying the OS. From left to right are the scatterplot of applying Zhang’s algorithm to the set 1, set 2, and set 3 of the QT database.

• 

  Figure 5. Comparison of the estimated T-end with or without optimal shrinkage on P-QRS-T cycles with high signal quality in the QT database. The black line is the preprocessed ECG signals from the first channel of one subject, and the light blue line is the same signals with the optimal shrinkage applied, where the whole signal is shifted down by 0.5 to enhance visualization. The red vertical line indicates the T-end point estimated by the chosen algorithm on the original signal. The green vertical line indicates the T-end point by the chosen algorithm on the denoised signal. The deep blue vertical line indicates the time point of T-end annotations provided by Cardio 1. Zhang’s algorithm is applied for the T-end detection. The case index is in the title.

• 

  Figure 6. Illustration of cardiac cycles variation. The gray lines are 20 cardiac cycles from subject id sel30 in QT database. The red line is the median over the 20 cycles.

• 

  Figure 7. Comparison of the estimated T-end points with or without optimal shrinkage over Ultra-Long-Term ECG recordings Database with intermediate quality. In each subplot, the black line is the original ECG signal, and the blue line is the denoised ECG signal estimated by the optimal shrinkage. The red (green) vertical lines indicate the estimated T-end by the chosen detection algorithm on noisy signals without (with) OS applied. The deep blue vertical line indicates the annotated T-end provided by EPICARE Center. The chosen T-end algorithm in the first, second and third column is Zhang’s, Carlos’s, and Martinez’s algorithm respectively.

• 

  Figure 8. A preliminary comparison of ECG morphologies with different metrics for arrhythmic ECG. For each subplot, the blue line is the ECG contaminated by ARMA (1,1) noise, the orange line is the ECG after OS with RRI as the metric, the yellow line is the ECG after OS with l2 norm as the metric, the purple line is the original ECG, green vertical lines are the locations of R peaks in each PVC occurrence, and the magenta spots are the corresponding estimations of T-end locations by Zhang’s algorithm. The signal is from the first channel of case 106 in MITDB. From top to bottom subplot, the noise level is 20 dB, 10 dB, and 5 dB respectively.