Author + information
- Received November 1, 2016
- Revision received February 1, 2017
- Accepted February 22, 2017
- Published online July 17, 2017.
- John L. Sapp, MDa,∗ (, )
- Meir Bar-Tal, MScb,
- Adam J. Howes, MDa,
- Jonathan E. Toma, MDc,
- Ahmed El-Damaty, MDd,
- James W. Warren, BSce,
- Paul J. MacInnis, BSce,
- Shijie Zhou, MASce and
- B. Milan Horáček, PhDe
- aThe QEII Health Sciences Centre, Halifax, Nova Scotia, Canada
- bBiosense Webster, Haifa, Israel
- cSunnybrook Health Sciences Centre, Toronto, Ontario, Canada
- dCairo University, Cairo, Egypt
- eDalhousie University, Halifax, Nova Scotia, Canada
- ↵∗Address for correspondence:
Dr. John L. Sapp, Queen Elizabeth II Health Sciences Centre, Room 2501-F Halifax Infirmary, 1796 Summer Street, Halifax, Nova Scotia B3H 3A7, Canada.
Objectives The aim of this study was to develop rapid computational methods for identifying the site of origin of ventricular activation from the 12-lead electrocardiogram.
Background Catheter ablation of ventricular tachycardia in patients with structural heart disease frequently relies on a substrate-based approach, which may use pace mapping guided by body-surface electrocardiography to identify culprit exit sites.
Methods Patients undergoing ablation of scar-related VT (n = 38) had 12-lead electrocardiograms recorded during pacing at left ventricular endocardial sites (n = 1,012) identified on 3-dimensional electroanatomic maps and registered to a generic left ventricular endocardial surface divided into 16 segments and tessellated into 238 triangles; electrocardiographic data were reduced for each lead to 1 variable, consisting of QRS time integral. Two methods for estimating the origin of activation were developed: 1) a discrete method, estimating segment of activation origin using template matching; and 2) a continuous method, using population-based multiple linear regression to estimate triangle of activation origin. A variant of the latter method was derived, using patient-specific multiple linear regression.
Results The optimal QRS time integral included the first 120 ms of the QRS interval. The mean localization error of population-based regressions was 12 ± 8 mm. Patient-specific regressions can achieve localization accuracy better than 5 mm when at least 10 training-set pacing sites are used; this accuracy further increases with each added pacing site.
Conclusions Computational intraprocedure methods can automatically identify the segment and site of left ventricular activation using novel algorithms, with accuracy within <10 mm.
Recurrent ventricular tachycardia (VT) is a long-term risk for patients with myocardial scar, often mitigated with an implantable cardioverter-defibrillator. The adverse prognosis and symptoms associated with recurrent VT and implantable cardioverter-defibrillator shocks drive the continued need to find ways to effectively suppress VT (1). Radiofrequency catheter ablation has been demonstrated to be an effective treatment strategy for patients with recurrent VT (2–4), but this therapy still has challenges. Point-by-point activation and entrainment mapping is effective in aiding ablation of inducible, hemodynamically tolerated VT but is not applicable in the majority of patients; substrate-based techniques are frequently necessary. These approaches may target specific regions of ventricular scar (5,6), regions of local abnormal ventricular activity (7–9), discrete channels (10–13), or all areas with abnormal signal amplitude.
The 12-lead electrocardiogram (ECG) has long been demonstrated to permit regional localization of endocardial pacing sites in patients with and without organic heart disease (14,15); various algorithms for localizing the origin of activation have been developed from extensive data accumulated in human studies (16–22). With 1 notable exception (20), localization algorithms used discrete anatomic regions to subdivide the endocardial surface and then to identify the region of most likely origin of activation. This approach suffers from unsatisfactory resolution. Epicardial activation can be accurately localized using body-surface potential mapping supplemented with the inverse solution but requires specialized electrodes and is complex in the setting of a clinical procedure (23). The aim of the present study was to develop a method that uses, in a complementary fashion, both discrete and continuous localization. We used the 12-lead ECGs acquired during pacing from known left ventricular (LV) endocardial sites—defined by 3-dimensional (3D) electroanatomic mapping—to develop discrete localization on the basis of template matching and a continuous localization based on linear regressions using a generic model of the LV endocardial surface. Finally, we investigated how the localization accuracy of the latter method can be improved by using each patient’s own LV geometry defined by the electroanatomic mapping system.
Sequential consenting patients with recurrent scar-related VT undergoing catheter-ablation procedures were enrolled. An initial cohort of 18 patients underwent 120-lead body-surface potential mapping during a catheter ablation procedure, followed by 20 patients who had only the standard 12-lead ECG recorded during the procedure (Table 1). For the purposes of this study, we used the 12-lead electrocardiographic data corresponding to known pacing sites and analyzed them off-line. All participating patients gave written informed consent; the protocol was approved by the institutional research ethics board. Patient flow within the study is detailed in the Online Appendix.
Electrophysiologic study and ablation
The electrophysiologic study was performed using standard techniques. Briefly, VT was induced by programmed stimulation from the right ventricular apex or outflow tract, and the left ventricle was mapped via the retrograde aortic or transseptal approach. Intracardiac electrograms were digitized and stored (CardioLab, GE Healthcare, Little Chalfont, United Kingdom), and 3D electroanatomic maps were acquired by the Carto XP or Carto 3 system (Biosense Webster, Diamond Bar, California). Substrate-based mapping and pace mapping were used to identify scar and potential culprit sites within the scar, which were targeted for ablation. An irrigated ablation catheter (Navistar Thermocool, Biosense Webster) was used for mapping and ablation. A complete LV endocardial electroanatomic map was created.
Data acquisition and processing
ECGs were recorded during pacing, and the 3D location of each pacing site was recovered from the electroanatomic mapping system. For the 120-lead electrocardiographic recordings, disposable strips with electrodes (FoxMed, Idstein, Germany) were used, and an acquisition system (Mark 6, BioSemi, Amsterdam, the Netherlands) with a computer running custom software displayed and recorded ECGs for 15 s (24). The 12-lead ECGs were acquired via analog output of the multichannel recording system (CardioLab), sampled at 1,000 Hz with 16-bit resolution and recorded for 15 s. Electrocardiographic recordings were reviewed off-line, a representative paced beat for each pacing site was selected, and the following features were calculated: 1) the time integral of the QRS complex (QRS-Int), a time integral of the entire QRS complex, in microvolt-seconds; 2) trimmed QRS-Int calculated over the initial 40, 50,…, 160 ms of the QRS complex; and 3) QRS-Int pattern, an ordered set of QRS-Int and/or trimmed QRS-Int values for the standard 12-lead ECG.
Generic model of the LV endocardium
A necropsy specimen of the normal human heart was used to construct the generic LV endocardial surface consisting of planar triangles in 3D space (25). Three coordinates specify each triangle’s vertices, in the orthogonal coordinate system with the origin at the LV apex and the z-axis oriented toward the midpoint of the aortic valve. A polar projection of this LV model comprising 238 triangles in 16 anatomic segments is shown in Figure 1.
Discrete localization by template matching
The 3D electroanatomic maps (sampling density <1.5 cm between points) acquired by the Carto system during the ablation procedure were reviewed, and each pacing site was associated with 1 of the 238 triangles of the generic LV model. Pacing sites within each of the 16 anatomic segments were used to generate QRS-Int templates, and QRS-Int patterns of each pacing site were then matched against each of the templates. For a QRS-Int pattern associated with a pacing site, the 16 segments were ranked according to the match, in terms of correlation coefficient, of their templates with the given pattern. The ability to localize the segment of origin was assessed by the percentage of matches ranked as first, second, or third most likely among 16 possible segments (“podium finish”).
Continuous localization by regression method
With the LV endocardium represented by the generic model, the origin of the coordinate system was translated to the center of the LV cavity, and using a training set of all available pacing sites with their corresponding ECGs, a regression model was fitted to this heart geometry in terms of electrocardiographic predictors (QRS integrals from the independent leads I, II, and V1 to V6), thus providing a statistical estimate of the 3D location of any activation site of interest on the basis of population-derived regression coefficients (see the Online Appendix for detailed methods).
The same linear regression approach can be applied to the patient’s own LV geometry, using individually generated regression coefficients, calculated from recorded ECGs while pacing at known sites. To minimize the number of required pacing sites, an exhaustive search for the optimal 3 leads was undertaken (see the Online Appendix for details). For 3-variable regressions, patient-specific regression coefficients can be calculated with data acquired by pacing at least 5 sites, although accuracy increases with more pacing sites. These calculations can be implemented in real time using a sweep-operator algorithm.
Prediction accuracy and localization performance of the regression method
Once the regression coefficients that best fit the training-set data are found, they can be applied as constants with the variables generated from the 12-lead ECG of any activation sequence of interest initiated at an unknown site. Thus, the 3D coordinates of an unknown pacing site or site of origin of ventricular arrhythmias can be calculated from known regression coefficients (either population based or patient specific) and the QRS-Int electrocardiographic variables for the unknown pacing or activation site of interest. If the actual pacing site is known in terms of its coordinates, the accuracy of pacing-site localization can be assessed from the distance between the actual and estimated site on the endocardial surface. The bootstrap method with replacement was used to assess the expected localization performance of the regression method by generating 1,000 random samples, each of the same size as the original one (n = 1,012), and the mean, median, and SD of localization error were calculated from all 1,000 trials.
Emulation of clinical protocols
Implementation of the patient-specific regression method with 3 electrocardiographic predictors requires that for at least 5 pacing sites, electrocardiographic variables together with coordinates of pacing sites be known. The sweep-operator algorithm can then calculate patient-specific regression coefficients for this minimal training set and, as more pacing sites are added to the training set, recalculate them to improve localization accuracy with each added pacing site. This process can be emulated off-line to assess its expected performance in real time, as demonstrated in the “Results” section.
All 3 methods were implemented for online applications on hand-held devices (Sony S tablet and Samsung Galaxy 3 phone) running Android 4.0. The patient’s 12-lead ECG is received via wireless connection, and the QRS onset is automatically identified; the user can move it if correction is necessary.
Thirty-eight patients undergoing catheter ablation for ventricular arrhythmias were enrolled (Table 1); the mean age was 62 ± 14 years; 37 patients were men; 77% had ischemic heart disease, 58% had undergone previous coronary bypass graft surgery, 89% had clinical VT, and the remaining patients had multiple premature ventricular complexes; and the mean ejection fraction was 35%. Electrocardiographic data corresponding to known LV pacing sites for all patients were pooled; after removing all recordings with stimulus-to-QRS delay >40 ms, a set of 1,012 known pacing sites with their corresponding 12-lead ECGs remained, which were used to investigate template matching and the population-based regression method. Among Patients #19 to #38, 790 pacing sites were analyzed to optimize the patient-specific regression method.
Registration of pacing sites
Three observers (A.E.-D., J.E.T., A.J.H.) independently registered pacing sites from the 3D Carto image by using anatomic landmarks or visual estimation on the basis of anatomic features, blinded to sites of automatic localization. They first demarcated 16 ventricular segments on the image and then selected the most appropriate triangle of the LV model (Figure 1) for each pacing site. Discrepancies were arbitrated in blinded fashion (J.L.S.).
Selection of optimal variables
To find optimal variables for representing electrocardiographic data, the performance of pacing-site localization by template matching and by the population-based regression method was tested for varied intervals of QRS-Int derived from 8 independent leads of the 12-lead ECG (Online Table 1). The optimal interval of integration for the QRS-Int variables included the first 120 ms of the QRS complex. For these optimal electrocardiographic input variables, template matching reached the best “podium-finish” results, and the smallest localization error was achieved by the population-based regression method.
Templates for segmental localization
Graphical representation of QRS-Int templates for 16 anatomic segments of the LV model is shown in Figure 2. Each bar represents the mean QRS-Int value over 120 ms from QRS onset of its corresponding lead. Each segment is associated with a characteristic QRS morphology.
Real-time localization using the template and regression methods
A representative recording in Figure 3A captures monomorphic VT shortly after it was induced. Corresponding displays of this recording, illustrating an implementation of both template matching and the regression method (using population-derived coefficients) on a handheld device, are shown in Figure 3B. The onset of 1 VT cycle was automatically detected and template matching identified the apical anterolateral segment #16 as containing the VT exit site; the regression method identified triangle #2 in the neighboring segment #15 (apical inferolateral) as the VT exit site. This correlated with the same triangle as the site of exit of the VT identified by activation and entrainment mapping (Figure 3C, yellow arrow).
Patient-specific localization using carto geometry of LV endocardium
An exhaustive search for the optimal electrocardiographic leads for calculation of patient-specific regression coefficients yielded leads III, V2, and V6 (see the Online Appendix for details). Tables 2 and 3 show how well the origin of activation can be localized, given 3 optimal electrocardiographic variables together with location coordinates for 10 pacing sites in the vicinity of the estimated origin of activation (obtained, for example, by using the population-based regression coefficients). Table 2 shows, for each patient separately, the accuracy of pacing-site localization that can be expected from 3-predictor linear regressions for optimal QRS-Int derived from the 3 optimal leads. Among 790 pacing sites, 131 were selected as “target” sites for testing the accuracy of pacing-site localization. For each patient, “target” sites suitable for testing patient-specific localization were identified as those that had at least 10 neighbors in the vicinity. Location coordinates of these 10 neighbors, together with 3 electrocardiographic predictors, were used to calculate patient-specific regression coefficients, and these were then used, in turn, to predict the location of the “target” site; the localization error of this prediction was calculated for each patient as mean, SD, and median of geodesic distance (in millimeters). From Table 3, the accuracy of pacing-site localization for patients who had target sites with neighboring pacing sites within 25-mm radius ranges from a mean of 2.6 to 8.2 mm (with 8 of 10 patients having a mean error <6 mm); for patients whose target sites had neighbors within a larger distance, mean error ranged from 4.5 to 20.7 mm (with 7 of 10 patients having mean error <9 mm). Table 3 provides another look at subsets of all 790 pacing sites. These pooled results from all patients indicate that for 37 target sites with the training set of 10 neighboring sites within an 18-mm radius, the mean error with standard deviation is 5.5 ± 4.7 mm; as evident from Figure 4, accuracy increased with increasing numbers of pacing sites within an 18-mm radius of the target site. For training-set pacing sites more remote from the target site, the mean error of localization moderately increases (Table 3).
Emulation of clinical work flow
A proposed work flow illustrates how patient-specific localization could be incorporated into a mapping and ablation procedure (Figure 5) and is illustrated for 2 patients (Figures 3 and 6). In these 2 cases, the 3D coordinates of VT exit sites were identified during the mapping and ablation procedure, on the basis of activation and entrainment mapping; 3-variable regressions were performed using the optimal ECG leads.
Representative Patient #1 (Table 1, Patient #27) had a prior inferior myocardial infarction with recurrent drug-refractory VT and underwent catheter ablation. One of the 4 episodes of inducible VT was mappable and localized using activation and entrainment mapping techniques and terminated with radiofrequency energy application at the apical inferolateral left ventricle. VT had a cycle length of 390 ms, with extreme axis deviation, and right bundle branch block–type configuration in lead V1 (Figure 3A). Using template matching, this VT was immediately (automatically) localized to the apical anterolateral segment as the first choice and the apical inferolateral segment as the second choice (Figures 2, 3A, and 3B). Using the regression method with population-based coefficients, it was localized at the junction of these 2 segments, identified by the “bull’s-eye” (Figure 3B). By directing the catheter to the region of interest and pacing at >5 sites within the region, patient-specific regression coefficients can be generated, which can be used to estimate the 3D coordinates of the site of ventricular activation. This becomes increasingly accurate with the addition of further pacing sites. Pacing from sites that result in stimulus-QRS delay >40 ms likely represent capture of an isthmus of tissue within myocardial scar, such that the site of myocardial “breakout” may not be the same as the pacing site; these sites should not be used for calculation of patient-specific regression coefficients. The confidence in the site of estimation increases with the addition of further pacing sites, achieving reproducibility within 5 mm after more than 11 pacing sites are included (Online Table 2). The site of VT exit identified by contact mapping in this case matches the estimated site in Figure 3C.
Representative Patient #2 (Table 1, Patient #37) had a history of prior myocardial infarction, prior VT treated with amiodarone, and an implantable cardioverter-defibrillator and presented with VT storm and underwent catheter ablation. One of the 5 episodes of inducible VT is illustrated in Figure 6A. This VT was inducible with double ventricular extrastimulation and had a cycle length 515 ms, with right bundle branch block–type configuration in lead V1 and a leftward axis. The site of exit was identified by activation and entrainment mapping prior to termination with radiofrequency application. Using (population-based) template matching, this was immediately and automatically localized to the basal inferior left ventricle as the first choice, with the second choice the basal inferolateral segment (Figure 6B). Using population-based coefficients and the regression method, it was localized in real time to the basal-lateral aspect of the inferobasal segment (Figure 6B, “bull’s-eye”). The electroanatomic map (Figure 6C, similar orientation as Figure 3C) with site of VT exit identified (yellow arrow and blue dot on mesh image) and estimated site of VT (purple dot, mesh image) are illustrated.
The purpose of this study was to investigate the ability of the 12-lead ECG to predict the site of origin of LV activation. The utility of the 12-lead ECG in localizing the origin of ventricular activation has been studied previously by Miller et al. (16,18) and by Kuchar et al. (17); their algorithms were limited to patients with anterior and/or inferior infarction, and the site of VT origin was determined with relatively coarse resolution. Segal et al. (21) described an algorithm for predicting LV endocardial exit site of scar-related VT that did not require knowledge of scar location, achieved high predictive accuracy, but classified the left ventricle into only 9 segments. Similarly, a more recent study by Yokokawa et al. (22) determined scar-related VT exit sites from the 12-lead ECG without reference to infarct location or other parameters but resolved the left ventricle endocardial surface to only 10 segments.
In our study, we initially investigated a discrete localization method of template matching that attempts to localize the origin of LV activation by using, as in the previous studies, relatively large endocardial segments. We have shown that QRS-Int variables derived from the 12-lead ECG can reliably localize the origin of ventricular activation with reasonable accuracy without prior knowledge of infarct-scar location. Our finding that the initial 120 ms of the QRS offers the highest localization accuracy to the segment of origin might indicate that the QRS waveform during the initial 120-ms interval is less dependent than the entire QRS complex on scar-tissue substrate.
An entirely different approach, previously attempted by Potse et al. (20) is a continuous localization method. Our proposed method of continuous localization estimates the origin of ventricular activation by means of linear regression that relates a generic 3D LV geometry to the electrocardiographic input variables (predictors). Regression equations use either general “population” coefficients calculated off-line from the large training set of ECGs for known pacing sites or patient-specific coefficients calculated during the electrophysiologic procedure. The latter method uses a pace-mapping approach to incrementally increase the accuracy of localization as the procedure continues, eventually localizing the origin of endocardial ventricular activation with mean accuracy of 5 mm. There are multiple ways this proposed patient-specific localization method could be incorporated into a VT ablation procedure. We propose that one could, after the creation of a 3D LV substrate map and VT induction, pace at 8 to 10 sites in the vicinity of the VT exit estimated by our population model, using these data to refine the localization estimate with the patient-specific regression method (Figure 5). Thus, the proposed patient-specific method of VT exit localization could simplify substrate ablation for VT and help increase the efficiency and possibly efficacy of catheter ablation procedures.
All pacing sites used to derive and validate both the template matching and the regression methods were from the LV endocardium. Thus, the applicability of the proposed techniques to the right ventricular endocardium, to the epicardial surface, and in the presence of individual clinical circumstances is yet to be established. Furthermore, the potential clinical utility of this technique will need to be established prospectively.
The site of initiation of LV activation can be automatically identified during a catheter ablation procedure by the proposed population-based localization methods with accuracy of 12 ± 8 mm using the standard 12-lead ECG. The patient-specific regression method of localization of ventricular activation can reduce the localization error to about 5 mm; this method uses the 12-lead ECG reduced to variables derived from only 3 electrocardiographic leads and becomes increasingly accurate as individual patient data are acquired by pacing from sites proximate to the site of origin of ventricular activation.
COMPETENCY IN MEDICAL KNOWLEDGE: Catheter ablation of the myocardial substrate responsible for scar-related VT may be enhanced by targeted ablation delivered at sites proximal to scar exit sites. Rapid automatic interpretation of the 12-lead ECG can be performed using computerized algorithms, possibly facilitating catheter ablation procedures.
TRANSLATIONAL OUTLOOK: Analysis of the surface ECG can localize sites of myocardial activation using both discrete (template matching) and continuous (regression method) techniques. The use of individualized regression coefficients derived from a limited number of pacing sites enhances the accuracy of localization. Further research should provide prospective clinical validation of this methodology.
This work was supported by grants from the Canadian Institutes of Health Research, the Heart & Stroke Foundation of Nova Scotia, and Biosense Webster. Dr. Sapp has received research grants from Biosense Webster, St. Jude Medical, and Philips Healthcare; and has received speaking honoraria from Medtronic and St. Jude Medical. Mr. Bar-Tal is an employee of Biosense Webster. Dr. Horáček has received research funding from Biosense Webster and Philips Healthcare. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
All authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the JACC: Clinical Electrophysiology author instructions page.
- Abbreviations and Acronyms
- left ventricular
- time integral of the QRS complex
- ventricular tachycardia
- Received November 1, 2016.
- Revision received February 1, 2017.
- Accepted February 22, 2017.
- 2017 American College of Cardiology Foundation
- Sapp J.L.,
- Wells G.A.,
- Parkash R.,
- et al.
- Soejima K.,
- Stevenson W.G.,
- Maisel W.H.,
- Sapp J.L.,
- Epstein L.M.
- Stevenson W.G.
- Harada T.,
- Stevenson W.G.,
- Kocovic D.Z.,
- Friedman P.L.
- Bogun F.,
- Good E.,
- Reich S.,
- et al.
- Jais P.,
- Maury P.,
- Khairy P.,
- et al.
- Arenal A.,
- del Castillo S.,
- Gonzalez-Torrecilla E.,
- et al.
- Tung R.,
- Mathuria N.,
- Michowitz Y.,
- et al.
- Jamil-Copley S.,
- Vergara P.,
- Carbucicchio C.,
- et al.
- Santangeli P.,
- Marchlinski F.E.
- Miller J.M.,
- Marchlinski F.E.,
- Buxton A.E.,
- Josephson M.E.
- Kuchar D.L.,
- Ruskin J.N.,
- Garan H.
- Sippensgroenewegen A.,
- Spekhorst H.,
- van Hemel N.M.,
- et al.
- Sapp J.L.,
- Dawoud F.,
- Clements J.C.,
- Horáček B.M.