Automated Localization of Focal Ventricular Tachycardia From Simulated Implanted Device Electrograms: A Combined Physics–AI Approach

Figures

• 

  FIGURE 1. Study workflow. The 3D torso model, generated (step 1) from a CT TAVI planning scan, was utilized to pace the LV from different locations, within a fast RE environment (step 2). These solutions were combined with LF matrices computed (step 3) on the 12-lead ECG electrodes and manufacture-guided CIED leads to generate accurate ECG and EGM traces (step 4), which were then used as inputs to two CNN architectures, one Cartesian probability-based (step 5), and one novel UVC-based (step 6). Localization of the paced beats across the LV was then computed (step 7) and compared to the actual locations of the simulated beats.

• 

  FIGURE 2. Patient-specific LV segment models. Generic AHA 17-segment model is shown in (F). The equivalent patient-specific model of the LV mesh is shown in (A) with basal, mid, and apical segments illustrated in (B–E), respectively. An example of the novel 68-segment model is shown in (H) highlighting the equal division in four parts of each of the 17 segments. (G) shows an example of how segment 1 in our mesh was divided into four equal segments.

• 

  FIGURE 3. Example of our simulation pipeline. RE single point excitations were simulated in ∼3767 different locations across the LV [an example of the activation time map generated for a single such paced beat is shown here in (A)]. These solutions were combined with the LF matrices computed on the standard nine ECG leads (V1–V6, RA, LA, and LL), and nine EGM sensing points (LVtip1–4, RVtip, RVring, RVcoil, SVC, and CAN = LA), here shown in (B) for LF on V1. The final signals at each lead, shown in (C), were then combined to obtain vector combinations shown in Figures 4E,F.

• 

  FIGURE 4. Torso setup. ECG and CIED configurations are shown in (A–D). Example of 16 combinations of pacing signals used for training and testing are illustrated in (E) for ECGs and (F) for EGMs.

• 

  FIGURE 5. Patient-specific bull’s-eye diagram. 3D patient-specific 17-segment model in (A,B) can be related to the 2D representation in (C) by considering UVC coordinates φ and z. Specifically, z (the distance from apex to base) can be linked to the radius of the 2D diagram, as shown, separately for each apical (Za), mid (Zm), and basal (Zb) part of the model. φ is linked to the segments along the spherical axis. The final patient-specific 2D 17-segment model can be seen in (D), where the various segments (1–17) are illustrated from blue-to-red color range [mapped from panel B].

• 

  FIGURE 6. UVC-based convolutional neural networks (CNNs). Structure of 68-feature classification (top) and 2-output regression (bottom).

• 

  FIGURE 7. Variations in ECG electrode placements. ECG leads were displaced by ∼5 cm upward (A), downward (B), toward the left (C), toward the right (D), and by ∼10 cm (mixed displacements toward the right and left) (E).

• 

  FIGURE 8. Variations in CIED configurations. Different CIED LV configurations according to different manufacturers (Boston Scientific, Livanova, Biotronik, and Medtronic) are shown on the top panels. The standard configuration of RV and LV leads is illustrated on the bottom left, and the septal RV coil configuration is shown on the bottom right. The main difference between the various configurations is the inter-electrode distance (reported above the manufacturers’ names).

• 

  FIGURE 9. Cartesian probability-based CNN performance. Testing performance of 17-feature Segment CNN is here reported in terms of accuracy (%) (A), and precision (B) for each of the 17 segments of the LV mesh. ECG-based and EGM-based testing performances are reported in blue and red, respectively.

• 

  FIGURE 10. Localization performance of Cartesian probability-based algorithm. Localization errors in millimeters are reported for ECG-based (blue) and EGM-based (red) testing. A comparison in localization performance between different Segment CNNs can also be seen; the 17-feature Segment CNN is on the left (A) and the 68-feature Segment CNN is on the right (B).

• 

  FIGURE 11. Localization comparison between UVC-based and Cartesian probability-based algorithms. Mean localization errors (A) are reported in millimeters with corresponding standard deviations for ECG-based (blue) and EGM-based (red) testing. An example of how ventricular tachycardia (VT) focal origins compare to UVC-based localized sources is shown in (B); diamonds represent the ground truths, whereas the circles are the CNN outputs. The gray color bar represents the distance from endocardial (black) to epicardial (white) surfaces of each source, whereas the blue-to-red color bar represents 1–17 patient-specific AHA segments.

• 

  FIGURE 12. Localization comparison between UVC-based (light gray) and Cartesian probability-based (dark gray) networks. Localization performance for ECG-based and EGM-based testing are here reported during sensitivity analysis for different noise levels—(A,B), respectively—and for different lead configurations—(C,D), respectively. For all panels, mean errors with corresponding standard deviations are reported for UVC-based (light gray) and Cartesian probability-based (dark gray) localization. Displacements of ECG leads, shown in Figure 7, results in localization errors in (C); on the other hand, different CIED configurations, shown in Figure 8, return errors in (D). Little difference in localization is present between the two networks with similar mean localization errors.