One population-based study (Dukes et al., 2015) of 1,139 older adults without any heart-failure signs or systolic dysfunction shows that premature ventricular complexes (PVC) and ventricular tachycardia (VT) burden are significantly associated with an increased risk of adjusted decreased left ventricular ejection fraction (odds ratio, 1.13) and increased adjusted risk of incident heart failure (hazard ratio, 1.06) and death (hazard ratio, 1.04). Catheter ablation (CA) is a commonly considered treatment of VT patients with and without structural heart disease when drugs are ineffective or have unacceptable side effects (Cronin et al., 2019). It has a class I indication for treatment of idiopathic outflow tract ventricular tachycardia (OTVT) (Joshi and Wilber, 2005; Latchamsetty et al., 2015). The OTVT stems from the right ventricular outflow tract (RVOT) in 60–80% of the cases and from the left ventricular outflow tract (LVOT) (Bunch and Day, 2006) in the rest of the cases. An accurate prediction of RVOT and LVOT origins of OTVT can optimize the CA strategy, reduce ablation duration, and avoid operative complications. Previous studies (Kamakura et al., 1998; Hachiya et al., 2000; Ito et al., 2003; Joshi and Wilber, 2005; Tanner et al., 2005; Haqqani et al., 2009; Zhang et al., 2009; Betensky et al., 2011; Yoshida et al., 2011, 2014; Cheng et al., 2013, 2018; Nakano et al., 2014; Efimova et al., 2015; He et al., 2018; Xie et al., 2018; Di et al., 2019; Enriquez et al., 2019; Yamada, 2019) propose several criteria or models to estimate RVOT and LVOT origins. However, these results have been limited by sample size, scope of studies, ECG measurement efficiency, and generalizability of the models. In contrast, we develop an optimal multistage scheme that automatically extracts features from standard 12-lead ECGs and incorporates these features into a machine learning model to predict RVOT and LVOT origins of VT or PVC with clinical-grade precision and provides multiprospective analysis for the most important ECG features.

We analyzed data from 420 patients who underwent CA of OTVT at the Ningbo First Hospital of Zhejiang University from March 2007 to September 2019. After the CA procedure, two (0.5%) patients developed slight ecchymosis. A total of five (1.2%) patients were excluded from this study because of frequent PVC or VT recurrence in the first 6-month follow-up.

Patient demographic and clinical characteristics data for the RVOT and LVOT groups are shown in Table 1. We compare the distributions of these background characteristics in the RVOT and LVOT groups and list the associated p-values in the table. The RVOT cohort consists of 20.95% left cusp, 17.62% posterior septal, 14.29% anterior septal, 10% anterior cusp, 7.86% free wall, and 7.14% right cusp. The LVOT cohort consists of 10.71% left coronary cusp, 5.71% aortomitral continuity, 2.62% left coronary cusp and right coronary cusp ommisure, 1.67% right coronary cusp, and 1.43% summit (shown in Supplementary Section A and Table 1).

The patients were assigned to training, validation, and testing cohorts, consisting of 340 (81%), 38 (9%), and 42 (10%) patients, respectively, using random proportional allocation (demographic summary shown in Table 1). For a fair comparison, the machine learning model was supplied with different features from two feature extraction methods. The performance was assessed using the same training, validation, and testing cohorts.

We used 1,600,800 automatically generated ECG features as machine learning model input. The proposed approach achieved an ACC of 97.62 (87.44–99.99); F₁-score of 98.46 (90–100); prediction of RVOT origins with SE of 96.97 (82.54–99.89); and SP of 100 (62.97–100) (shown in Table 2), respectively; and AUC of 98.99 (96.89–100) (presented in Figure 5). Among the 1,600,800 initial automatically generated ECG features, we found a total of 1,352 critically important features with non-zero Shapley additive explanations (SHAP) values (Lundberg and Lee, 2017), showing the importance of their contributions to RVOT and LVOT prediction. The detailed interpretation of SHAP value is introduced in Supplementary Section C.1. We chose and analyzed the top three important features (shown in Figure 6) that have significant classification capability: (Dukes et al., 2015) the ratio between the location of the 5th peak or valley at the SR beat V₁ lead and the right boundary of the 5th peak or valley at the V₁ lead of PVC, Cronin et al. (2019) the ratio between the prominence of the 5th peak or valley at the V1 lead of PVC and the prominence of the 5th peak or valley at the V3 lead of PVC, and (Joshi and Wilber, 2005) the difference between the distance of the 5th peak or valley to the left boundary at the V1 lead of PVC and the distance of the 5th peak or valley to the left boundary at the V₁ lead of the SR beat.

Training the machine learning model using 155,784 features extracted from conventional QRS morphological ECG measurements, the proposed method attained an ACC of 92.86 (80.35–98.85), F₁-score of 95.38 (86.62–98.86), prediction of RVOT origins with SE of 93.94 (78.64–98.99) and SP of 88.89 (50.86–99.45) (shown in Table 2), and AUC of 95.62 (89.78–100) (presented in Figure 5). Among the initial 155,784 features, we found a total of 1,003 critically important features with non-zero SHAP values (Lundberg and Lee, 2017), showing the importance of their contributions to RVOT and LVOT prediction. The top three important features (shown in Supplementary Section C.1 and Figure 2) that show significant classification capability are (Dukes et al., 2015) the ratio between R-wave amplitude based on the zero isoelectric baselines at lead III PVC and the R-wave amplitude based on the offset of S-wave at V₁ lead PVC, Cronin et al. (2019) the ratio between the R-wave amplitude based on R-wave onset at V₂ lead SR beat and the R-wave amplitude based on zero isoelectric baseline at V₃ lead PVC, and (Joshi and Wilber, 2005) the ratio between the R-wave amplitude based on the zero isoelectric baseline at aVL lead SR beat and the R-wave amplitude based on S-wave offset at V₁ lead PVC. The statistical summary of conventional QRS morphological measurements for leads V₁ to V₆ is listed in Supplementary Section A and Table 2.

Finally, the average performance of eight cardiologists who determined RVOT and LVOT using the same ECG samples in this study is presented in Table 2. The classification confusion matrix for these three methods shows correct and incorrect frequency counts in Supplementary Section A and Table 3. Furthermore, we compared our approach against related methods from 12 prior studies (Kamakura et al., 1998; Zhang et al., 2009; Betensky et al., 2011; Yoshida et al., 2011, 2014; Cheng et al., 2013, 2018; Nakano et al., 2014; Efimova et al., 2015; He et al., 2018; Xie et al., 2018; Di et al., 2019). ACC, F₁-score, SE, SP, positive predictive value, negative predictive value, and AUC were used to compare performances and are shown in Table 3.

We designed and implemented a high-accuracy algorithm for LVOT and RVOT origins of OTVT classification, using 1,600,800 ECG measurements automatically extracted from 12-lead ECGs using our proprietary method. The prediction accuracy comparison among our method combined with the XGBoost classifier, a conventional QRS feature extraction method combined with XGBoost, and the performance of human experts (shown in Table 2) shows that the machine learning model with the automated ECG feature extraction method was uniformly superior. We used DeLong’s test (DeLong et al., 1988) to demonstrate that the automated ECG feature extraction method had a significantly higher AUC compared with that attained by the conventional QRS morphological feature extraction approach with a P-value = 0.035. The comparison of our approach against methods from 12 prior studies (Kamakura et al., 1998; Zhang et al., 2009; Betensky et al., 2011; Yoshida et al., 2011, 2014; Cheng et al., 2013, 2018; Nakano et al., 2014; Efimova et al., 2015; He et al., 2018; Xie et al., 2018; Di et al., 2019) shows that our algorithm achieved the highest performance scores (shown in Table 3). Additionally, we evaluated the general classification capability of each criterion proposed by previous studies using the database in this study. Not surprisingly, we observed significant differences between previously reported performances and the reproduced results of these methods because most of the prior studies used the univariate analysis to make predictions (shown in Table 3).

The excellent performance of our machine learning algorithm demands an enormous volume of data and features. It is an extremely time- and cost-consuming task to generate such amount of features by the conventional ECG QRS morphological measurements introduced in prior studies because these measurements are manually obtained. Thus, we did not make any assumptions about ECG criteria before training the machine learning algorithm and intended to exhaust all possible relationships among morphological measures of Q-, R-, and S-waves as well as the entire QRS complex. We designed and implemented an automated ECG feature extraction method that can generate 1,600,800 ECG signal characteristics. Not only did these features contain a considerable amount of the classical statistics from 12 prior studies (Kamakura et al., 1998; Zhang et al., 2009; Betensky et al., 2011; Yoshida et al., 2011, 2014; Cheng et al., 2013, 2018; Nakano et al., 2014; Efimova et al., 2015; He et al., 2018; Xie et al., 2018; Di et al., 2019), but they also captured morphological measures not considered by previous studies, such as rsR’ waves and rsr’s’ waves. However, one may be concerned that such a feature extraction method will include the P- and T-wave within SR beats and retrograde P-waves within PVC. The machine learning model captures and analyzes a large amount of information from every beat but filters out all unimportant features based on their classification accuracy contribution. As we can see from the top three important features (shown in Figure 6) selected by the machine learning model, none of the features that presented waves mentioned above played a role in the prediction. The important morphological features of the Rsr’ and rsr’s waves may be caused by noise and lead placement of the 12-lead ECG electrodes because the 12-lead ECG electrodes are frequently misplaced due to the mapping patches used during the ablation procedure. In this study, we avoid such a problem because chest and limb leads were placed carefully in a standard position when the 12-lead surface ECGs were collected before the procedure.

Moreover, before the machine learning model is safely applied in practice, an unambiguous interoperation is necessary for cardiologists to gear this advanced tool, such as explaining what crucial criteria are and why they play vital roles. For instance, the machine learning model shows that the smaller the magnitude of the first important feature (shown in Figure 6C.1), the higher the possibility of LVOT origin of OTVT. The first important feature is the ratio of the location of the 5th peak or valley at the V1 lead SR beat and the right boundary of the 5th peak or valley at the V₁ lead of PVC. In our feature extraction system, the 5th peak or valley at the V₁ lead of PVC is an S-wave in most cases. The key ECG lead in the initial site prediction of VT origin is the V₁ lead because it is located nearly orthogonal to the septal plane and, thus, is the best lead to resolve initial right- vs. left-sided activation. When the V₁ lead has a positive QRS (R > s), the VT is considered to have the right bundle branch block (RBBB) configuration. Conversely, net negative QRS (r < S) defines a left bundle branch block (LBBB) configuration (Haqqani and Marchlinski, 2019). The top three important features (shown in Figure 6) were exactly measured activation time, RBBB, and LBBB configuration. Therefore, such interpretation makes the machine learning decision process not a black box anymore.

Last but not least, the machine learning model proposed in this study can be immediately and effortlessly deployed to EP labs. The pretrained model, source code, and data are available online and found in the “Data Availability Statement” section. The model inputs are only two QRS complexes, one for PVC and one for SR beat, and they can be easily acquired from 12-lead standard ECG. The analysis of one patient’s data takes less than a second provided every step of measurement and computation is automatically done by the model and the preprocessing approach. The precise prediction of origins can significantly reduce CA duration and reduce the risk of complications.

Considering the performance of prediction, the capacity of extracting vital information from 12-lead ECG and the robustness of application, our results provide the promising and reliable decision support to guide a successful CA treatment of ventricular arrhythmia by machine learning technology.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://doi.org/10.6084/m9.figshare.c.4668086.v2.

JZ, GF, XD, BH, HC, and CR processed the data for analysis. JZ, HC, GF, XD, IA, and CR performed the statistical analysis. All authors contributed to the study design, data interpretation, and writing of the report.

From the Department of Cardiology, Ningbo First Hospital of Zhejiang University. HC has served as a consultant for Biosense Webster, Boston Scientific, and Abbott. HY was employed by the company Zhejiang Cachet Jetboom Medical Devices Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.