Automatic wide complex tachycardia differentiation using mathematically synthesized vectorcardiogram signals

Abstract Background Automated wide complex tachycardia (WCT) differentiation into ventricular tachycardia (VT) and supraventricular wide complex tachycardia (SWCT) may be accomplished using novel calculations that quantify the extent of mean electrical vector changes between the WCT and baseline electrocardiogram (ECG). At present, it is unknown whether quantifying mean electrical vector changes within three orthogonal vectorcardiogram (VCG) leads (X, Y, and Z leads) can improve automated VT and SWCT classification. Methods A derivation cohort of paired WCT and baseline ECGs was used to derive five logistic regression models: (i) one novel WCT differentiation model (i.e., VCG Model), (ii) three previously developed WCT differentiation models (i.e., WCT Formula, VT Prediction Model, and WCT Formula II), and (iii) one “all‐inclusive” model (i.e., Hybrid Model). A separate validation cohort of paired WCT and baseline ECGs was used to trial and compare each model's performance. Results The VCG Model, composed of WCT QRS duration, baseline QRS duration, absolute change in QRS duration, X‐lead QRS amplitude change, Y‐lead QRS amplitude change, and Z‐lead QRS amplitude change, demonstrated effective WCT differentiation (area under the curve [AUC] 0.94) for the derivation cohort. For the validation cohort, the diagnostic performance of the VCG Model (AUC 0.94) was similar to that achieved by the WCT Formula (AUC 0.95), VT Prediction Model (AUC 0.91), WCT Formula II (AUC 0.94), and Hybrid Model (AUC 0.95). Conclusion Custom calculations derived from mathematically synthesized VCG signals may be used to formulate an effective means to differentiate WCTs automatically.


| INTRODUC TI ON
Twelve-lead electrocardiogram (ECG) interpretation is the most practical means to non-invasively differentiate wide complex tachycardias (WCTs) into ventricular tachycardia (VT) or supraventricular wide complex tachycardia (SWCT). Rigorous research spanning several decades has amassed an expansive arsenal of manual ECG interpretation methods , each relying upon the visual recognition of distinctive electrocardiographic features of VT and SWCT. Yet, despite the creation of numerous manual diagnostic criteria and algorithms, arriving at a correct and timely VT or SWCT diagnosis remains problematic.
Recent works May et al., 2019aMay et al., , 2020 have introduced several novel automated methods capable of distinguishing VT and SWCT with high accuracy. Through the use of readily accessible ECG data routinely processed by computerized ECG interpretation software, automated methods (i.e., WCT Formula [2019] (May et al., 2019a), VT Prediction Model [2020] , and WCT Formula II [2020] ) are able to deliver to clinicians an estimation of VT probability-one that is freely independent of ECG interpreter competency. By design, each automated approach makes use of paired WCT and baseline ECG data to deduce the magnitude of mean electrical vector changes contained within two ECG planes (i.e., frontal [limb leads] and horizontal [chest leads]). In the case of the WCT Formula (May et al., 2019a), the frontal and horizontal percent amplitude change (PAC) calculations are used to broadly quantify QRS amplitude (μV) changes in the frontal and horizontal planes, respectively.
Similarly, the WCT Formula II , comprised of frontal and horizontal percent timevoltage area change (PTVAC) calculations, makes use of QRS timevoltage area (TVA) (μV•ms) changes in the frontal and horizontal planes, respectively. Alternatively, the VT Prediction Model  uses QRS axis (°) and T-wave axis (°) change, both of which can be easily computed from standard computerized ECG measurements.
In this work, we sought to determine whether the quantification of QRS amplitude changes, between paired WCT and baseline ECGs, within three orthogonal vectorcardiogram (VCG) leads (i.e., X-lead [patient's right to patient's left], Y-lead [cranial-to-caudal], and Z-lead [anterior-to-posterior]), may yield effective methods to differentiate WCTs automatically. Provided that QRS complex data from mathematically synthesized X-, Y-, and Z-VCG leads can be configured to determine the mean electrical vector of depolarization across three spatial planes (i.e., frontal plane [X and Y leads], horizontal plane [X and Z leads], and sagittal plane [Y and Z leads]) (Figure 1), we hypothesized that QRS amplitude changes of mathematically synthesized VCG leads may offer a more robust means of quantifying changes in the mean electrical vector of depolarization and enable more accurate WCT differentiation. F I G U R E 1 Mean ventricular depolarization vector in the frontal, horizontal, and sagittal ECG planes. The X-lead appraises electrical changes from the patient's right to patient's left direction (frontal and horizontal ECG planes), the Y-lead appraises electrical changes from the cranial-to-caudal direction (frontal and sagittal ECG planes), and the Z-lead appraises electrical changes from the anterior-to-posterior direction (horizontal and sagittal ECG planes). Yellow arrows illustrate the archetypal direction and magnitude of a mean electrical vector of ventricular depolarization for a normal baseline ECG. The spatial orientation of each lead is depicted using blue (VCG leads) and black (standard 12-lead ECG leads) font lettering. The directionality of VCG signal recordings is depicted by a "+" symbol (positive voltage-i.e., waveforms above the isoelectric baseline). ECG, electrocardiogram; VCG, vectorcardiogram K E Y W O R D S electrocardiogram, supraventricular tachycardia, ventricular tachycardia, wide complex tachycardia 2 | ME THODS

| Study design
In this study, we formulated and trialed two novel logistic regression models (i.e., VCG Model and Hybrid Model) comprised of custom computations from mathematically synthesized VCG signals derived from paired WCT and subsequent baseline ECGs. The diagnostic performance of each model was directly compared to other previously described automated WCT differentiation models, namely the WCT Formula (May et al., 2019a), VT Prediction Model , and WCT Formula II  Patient data acquisition and analysis was approved by the Mayo Clinic Institutional Review Board. Clinical and electrocardiographic data from patients of the derivation and validation cohorts were previously examined and described in prior works May et al., 2019b). WCTs were required to satisfy standard WCT criteria (QRS duration ≥120 ms and ventricular rate ≥100 beats per minute) and possess an official ECG laboratory interpretation of (i) "ventricular tachycardia," (ii) "supraventricular tachycardia," or (iii) "wide complex tachycardia." Baseline ECGs were defined as the first non-WCT rhythm recorded after the WCT event. Polymorphic WCTs and WCTs demonstrating grossly irregular atrioventricular conduction (e.g., atrial fibrillation or atrial flutter with variable atrioventricular block) were excluded.

| Electrocardiogram selection
ECGs demonstrating truncated WCTs (e.g., brief run of non-sustained VT) occurring within a dominant baseline heart rhythm (e.g., normal sinus rhythm) were not evaluated. If a WCT did not have a baseline ECG or definitive clinical diagnosis established by the patient's overseeing physician, it was excluded from further analysis.

| Official ECG laboratory diagnosis
Official ECG interpretation was completed by expert ECG interpreters, including 7 heart rhythm cardiologists and 14 non-heart rhythm cardiologists.

| Clinical diagnoses
Clinical diagnoses (i.e., VT or SWCT) were established by the patient's supervising physician. Physicians responsible for clinical diagnoses were stratified according to a subjective hierarchy of clinical expertise: (i) heart rhythm cardiologist, (ii) non-heart rhythm cardiologist, and (iii) non-cardiologist. All physicians responsible for clinical diagnoses had access to the official ECG interpretation diagnosis provided by the ECG laboratory.

| ECG measurements
2.6.1 | Computerized ECG measurements Standard computerized ECG measurements for WCT and baseline ECGs, including QRS duration (ms), QRS axis (°), and T-wave axis (°), were generated by GE Healthcare's MUSE ECG interpretation software. Computerized QRS amplitude (μV) and TVA (time-voltage area) (μV•ms) measurements of waveforms above (r/R and r'/R') and below (q/QS, s/S, and s'/S′) the isoelectric baseline were automatically derived from the dominant QRS complex template of select ECG leads (i.e., aVR, aVL, aVF, V1, V4, and V6). Only amplitude and TVA measurements representative of QRS complex waveforms were analyzed.

| Manual VCG measurements
Three mathematically synthesized VCG signals (i.e., X, Y, and Z leads) were automatically generated by GE Healthcare's MUSE ECG interpretation software package. VCG signal QRS amplitude (μV) measurements of waveforms above (r/R and r'/R') and below (q/QS, s/S, and s'/S′) the isoelectric baseline were directly measured (using electronic calipers provided by the MUSE ECG interpretation software package) by the first author (K.A.H.), who was blinded to patients' clinical characteristics and final rhythm diagnosis (VT or SWCT). Special attention was made to exclude pacing stimuli (i.e., "pacing spikes") from QRS complex amplitude measurements.

| ECG parameters
2.7.1 | QRS duration change (ms) QRS duration change denotes the absolute difference in QRS duration (ms) measurements between paired WCT and baseline ECGs.

| QRS axis change (°)
QRS axis change is the absolute difference in the frontal plane QRS axis (°) between paired WCT and baseline ECGs.

| T-wave axis change (°)
T-wave axis change represents the absolute difference in the frontal plane T-wave axis (°) between paired WCT and baseline ECGs.
Unlike frontal and horizontal PAC or PTVAC, which can both be calculated from automated computerized QRS waveform measurements of standard ECG leads (aVR, aVL, aVF, V1, V4, and V6), X-, Y-, and Z-lead QRS amplitude change calculations are derived from manual QRS waveform measurements of mathematically synthesized VCG signals generated by computerized ECG interpretation software. The mathematical procedure necessary to calculate X-, Y-, and Z-lead QRS amplitude change is presented in Figure S5.

| Logistic regression models
2.9.1 | WCT formula, VT prediction model, and WCT formula II Prior work has described the logistic regression model structure of the WCT Formula (May et al., 2019a(May et al., , 2019b, VT Prediction Model May et al., 2020), and WCT Formula II . In general, each model uses independent VT predictors to derive an automatic estimation of VT probability (0.00%-99.99%). The logistic regression structure of the WCT Formula, VT Prediction Model, and WCT Formula II are shown in Figures S6-S8.

| VCG model
By design, the VCG Model integrates measured and calculated ECG data from paired WCT and baseline ECGs to deliver an unambiguous estimation of VT probability (0.00%-99.99%) using independent VT predictors synchronously weighted according to their influence on the binary classification of VT or SWCT. The logistic regression structure of the VCG Model is shown below: The VCG Model incorporates six parameters directly accessed or calculated from paired WCT and baseline ECG data: two standard ECG measurements (i.e., WCT QRS duration [ms] and baseline QRS X = β 0 + β 1 X 1 + β 2 X 2 + β 3 X 3 + β 4 X 4 + β 5 X 5 + β 6 X 6 = Ln P 1 − P X = −12.4466 + (0.009548) (X − lead QRS amplitude change) , one arithmetic calculation of standard ECG measurements (i.e., QRS duration change [ms]), and three novel computations derived from mathematically synthesized VCG signals (i.e., X-, Y-, and Z-lead QRS amplitude change [%]). Each of the six parameters (Xx) are apportioned a beta coefficient (βx) according to their influence on binary classification. The "constant" term (β 0 ) denotes the y-intercept for the least-squares regression line. The weighted sum predictor (X β ) and VT probability (P) may be calculated after integrating VT predictor (Xx) values derived from WCT and baseline ECG data.

| Statistical analysis
Categorical variables were compared using chi-square tests.
Wilcoxon rank-sum tests were used to compare continuous variables. In order to ensure satisfactory model comparisons, all logistic regression models were originally formulated or re-derived using the same collection of ECG pairs comprising the derivation cohort. Thereafter, each model was trialed separately on the same validation cohort. Previously described logistic regression models, including the WCT Formula, VT Prediction Model, and WCT Formula II, were composed of the same variables as defined by the original works introducing each model May et al., 2019aMay et al., , 2020. Paired ECGs of the validation cohort were assigned estimated VT probabilities (0.00%-99.99%) by all five logistic regression models. Binary rhythm classification (VT or SWCT) was rendered according to a pre-specified VT probability partition of 50% (i.e., VT ≥ 50% and SWCT < 50%).
Performance metrics (i.e., accuracy, sensitivity, specificity, positive predictive value [PPV], and negative predictive value [NPV]) for each model were assessed according to their agreement with the overseeing physician's clinical diagnosis. AUC was used to summarize overall diagnostic performance. Comparison of the fit between the statistical models was completed using a Delong test. A two-tailed p-value of <.05 was considered statistically significant. Statistical analyses were completed using SAS version 9.4 (SAS Institute).

| Derivation cohort
The derivation cohort included 400 paired WCT (185 VT, 215 SWCT) and baseline ECGs from 309 patients. Clinical diagnosis and ECG laboratory interpretation data are shown in Table S1. The majority (86.0%) of clinical diagnoses (VT or SWCT) were established by heart rhythm or non-heart rhythm cardiologists. Two-hundred and three out of the 400 (50.8%) VT or SWCT clinical diagnoses were established among patients having a corroborating electrophysiology procedure or implanted intra-cardiac device (e.g., pacemaker).
Patient characteristics of VT and SWCT groups are described in

| Validation of diagnostic performance
Overall accuracy, sensitivity, specificity, PPV, NPV, and AUC of each logistic regression model is listed in Table 2. Comparisons of diagnostic performance metrics between each logistic regression model are presented in Table 3.  Summary of logistic regression model performance metrics. Overall accuracy, sensitivity, specificity, positive predictive value, and negative predictive value rendered according to a pre-specified VT probability partition of 50% (i.e., VT ≥ 50% and SWCT < 50%). Numbers in parentheses are 95% confidence intervals.
Abbreviations: AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value.

TA B L E 3
Comparison of logistic regression model performance was similar whether or not the VCG Model was applied to patients who were diagnosed by the traditional "gold standard" methods, including rhythm classification based on the results of electrophysiology procedures or analysis of intra-cardiac device (i.e., pacemaker or ICD) recordings. The use of a pre-specified 50% VT probability cut-point for WCT classification (i.e., VT ≥ 50% and SWCT < 50%) yielded favorable diagnostic indices, including strong overall accuracy with high diagnostic sensitivity and specificity for VT.

| X-, Y-, and Z-lead QRS amplitude change
In prior works May et al., 2019a), we demonstrated that the quantification of QRS amplitude or TVA changes in specific leads of the frontal (aVR, aVL, aVF) and horizontal (V1, V4, V6) ECG planes enable accurate WCT differentiation. The underlying conceptual basis for these observations is that VT, which may originate and spread from any location within the right or left ventricles, categorically demonstrates immense "electrical freedom" compared to SWCT, which ordinarily depolarizes the ventricular myocardium in a manner prescribed by the native His-Purkinje network . Consequently, VT may express a virtually unlimited diversity of distinct QRS complexes dissimilar from those present on the respective baseline ECG.
Correspondingly, VT will commonly demonstrate greater changes to the electrical vector of depolarization than SWCT upon WCT onset or offset. The few notable exceptions to this concept include VTs that rapidly engage (e.g., fascicular VT) or characteristically make use of the heart's His-Purkinje network (e.g., bundle branch reentry).
In contrast, SWCTs are expected to ordinarily demonstrate a more constrained range of mean electrical vectors and limited number of electrocardiographically distinct QRS complexes compared to the patient's baseline ECG. In rarer circumstances, marked mean electrical vector and QRS complex changes may arise from SWCTs that develop from impulse propagation using atrioventricular accessory pathways (i.e., pre-excitation).
Similar to other previously introduced variables (e.g., frontal PAC, horizontal PTVAC, and QRS axis change) May et al., 2019aMay et al., , 2020, X-, Y-, and Zlead QRS amplitude change are designed to quantify the extent of mean electrical vector changes that occur upon WCT onset or offset. However, unlike previously described diagnostic variables, the collective evaluation of X-, Y-, and Z-lead QRS amplitude change enables the direct quantification of mean electrical vector changes occurring in three, instead of two, spatial planes (i.e., patient's right to patient's left [X-lead], cranial-to-caudal [Y-lead], and anterior-toposterior [Z-lead]) ( Figure 3). Therefore, we hypothesized X-, Y-, and F I G U R E 3 Mean electrical vector changes in the frontal, horizontal, and sagittal ECG planes. Summary of expected changes to the mean electrical vector of ventricular depolarization following WCT initiation within three orthogonal ECG planes (i.e., frontal, horizontal, and sagittal). Displayed arrows represent mean electrical vectors for ventricular depolarization. The directional orientation of individual arrows expresses the mean electrical axis of ventricular depolarization (i.e., QRS axis). The heavy yellow arrows denote the baseline heart rhythm's mean electrical vector for ventricular depolarization. Color-shaded regions and arrows denote the expected range of mean electrical vectors after WCT onset. The spatial orientation of VCG leads is depicted with blue font lettering. The directionality of ventricular depolarization captured by VCG leads is portrayed by a "+" symbol (positive voltage-i.e., waveforms above the isoelectric baseline). The X-lead appraises electrical changes from the patient's right to patient left direction (frontal and horizontal ECG planes), the Y-lead appraises electrical changes from the cranial-to-caudal direction (frontal and sagittal ECG planes), and the Z-lead appraises electrical changes from the anterior-toposterior direction (horizontal and sagittal ECG planes We report X-, Y-, and Z-lead QRS amplitude change are strong independent VT predictors-VT generally generates greater X-, Y-, and Z-lead QRS amplitude change than SWCT. As such, X-, Y-, and Despite what would be an incredible diagnostic advantage, a glaring limitation of the VCG Model, Hybrid Model, and other recently introduced automated approaches is that they require the use of a recorded and digitally archived baseline ECG for their application. In circumstances where WCT patients present without a baseline ECG, clinicians and ECG interpreters would have to temporarily rely on manual WCT differentiation methods until the patient's baseline ECG is acquired. In future works, the customized variables that make up the VCG Model and other automated WCT differentiation models may be integrated with additional computerized ECG measurements or yet to be formulated diagnostic determinants to bolster diagnostic performance. Additionally, customized variables, which comprise already described logistic regression models, could be used by more sophisticated modeling techniques (e.g., artificial neural networks) more apt to decipher meaningful non-linear and non-parametric relationships.

| Study limitations
Study limitations were comprehensively described in prior works May et al., 2019aMay et al., , 2020. However, there are two important limitations that warrant special attention. First, our study evaluated any clinically encountered WCT (i.e., "all-comers") that was formally diagnosed by the patient's overseeing physician. As a result, this analysis did not exclude patients who did not undergo an electrophysiology procedure or did not have an implanted intra-cardiac device (e.g., ICD or pacemaker). Although we found model performance did not differ among patients with or without an accompanying electrophysiology procedure or implanted intra-cardiac device, we must acknowledge that a substantial proportion of VT or SWCT diagnoses were not established by the traditional "gold standard." Furthermore, this WCT selection process precluded a detailed assessment of automated model performance among various SWCT (e.g., SWCT due to bystander conduction over various accessory pathways) and VT (e.g., fascicular VT) subtypes. Nevertheless, by not excluding WCT tracings lacking a corroborating electrophysiology procedure or implanted intra-cardiac device, our analysis counters the patient selection biases created by only evaluating a proportionally smaller subgroup of WCTs seen in clinical practice. Second, the diagnostic performance of automated models was not directly compared with traditional manual WCT differentiation approaches (Brugada et al., 1991;Chen et al., 2019;Jastrzebski et al., 2017;Pava et al., 2010;Vereckei et al., 2008). Additional research will be necessary to determine whether automated models exhibit diagnostic superiority over conventional manual interpretation methods. well when compared to other high-performing automated WCT differentiation models and may be incorporated into existing ECG interpretation software systems.

CO N FLI C T O F I NTE R E S T
This technology was disclosed to Mayo Clinic Ventures who possesses intellectual property rights.