Deep-Learning-Based Estimation of the Spatial QRS-T Angle from Reduced-Lead ECGs

Figures

• 

  Figure 1. Overview of the proposed deep learning model for estimation of QRS-T angle using reduced-lead ECGs. The model is composed of two parts: feature extraction and regression. The target vectors u→QRS and u→T and spatial QRS-T angle α are computed from VCGs.

• 

  Figure 2. Detailed representation of the three types of blocks employed in the feature extraction network: (a) first block, (b) the last block, and (c) blocks with residual connections.

• 

  Figure 3. Case scenarios of: (a) similar QRS-T angles α of two u→QRS and u→T located in two different planes; (b) correct location of one vector (u→T) but not the other (u→QRS), yielding large errors in the estimated QRS-T angle α^; (c) compromise between minor errors in the location of both u→QRS and u→T to achieve a more accurate QRS-T angle estimation.

• 

  Figure 4. Data preparation and labeling. Signals undergo preprocessing to generate the input signal-averaged beats.

• 

  Figure 5. Distribution of spatial QRS-T angles α of across the ranges of α = [0:5:180]° according to sex (overlapped) for all eligible recordings in the dataset (left) and for each morphological class (right). α is the angle between the VCG vectors u→QRS and u→T. The dashed line is the median α for each class.

• 

  Figure 6. Distribution of spatial QRS-T angle α across the ranges of α = [0:5:180]° according to sex (overlapped) for all recordings suitable for analysis (left) and for each morphological class (right) in the (a) training and (b) validation sets. The dashed line is the median α for each class.

• 

  Figure 7. Performance of various model configurations tuned with different combinations of hyperparameters w1 and w2 in estimating the spatial QRS-T angle from leads XYZ in the validation dataset. (a) Boxplot of the obtained absolute error ϵ (outliers not shown) for every combination of w1 and w2. w1 increases to the left side, wheres w2 to the right. The other hyperparameter value is obtained as |1−w| on each side. (b) Mean absolute error ϵ¯ and the respective 95% confidence interval across the ranges of α = [0:5:180]° for increasing w1 (top row) and w2 (bottom row). The last column displays the number of training samples for each range of α.

• 

  Figure 8. Scatter plot diagrams of the deep-learning-estimated α^ vs. target α from various sets of leads for (top row) all recordings, and ECGs with normal (middle row) NORM and (bottom row) cardiac disease in the validation dataset. The estimation error ϵ of every α^ in the first row is color-grouped based on the absolute median (ϵ˜), mean (ϵ¯), and standard deviation (σϵ) error.

• 

  Figure 9. Variation of the mean absolute error ϵ¯ and the respective the 95% confidence interval across the ranges of α = [0:5:180]° for ECGs with normal (NORM) and diseased cardiac function. The right axis indicates the number of training samples in each range of α. Since the number of NORM subjects with α > 120° is almost negligible, ϵ¯ is not shown for these ranges of α.

• 

  Figure 10. Bland-Altman diagrams of deep-learning-based estimation of α^ from various sets of leads of ECGs with (top) normal (NORM) and (bottom) diseased cardiac function.

• 

  Figure 11. Distribution of the Euclidean distance Ldu→,u→^ between u→QRS and u→^QRS, and u→T and u→^T in each of the three planes: XY (frontal), XZ (transverse), and YZ (sagittal).