Circulation Journal
Online ISSN : 1347-4820
Print ISSN : 1346-9843
ISSN-L : 1346-9843

This article has now been updated. Please use the final version.

Development of a Visualization Deep Learning Model for Classifying Origins of Ventricular Arrhythmias
Kazutaka NakasoneMakoto NishimoriKunihiko KiuchiMasakazu ShinoharaKoji FukuzawaMitsuru TakamiMustapha El HamritiPhilipp SommerJun SakaiToshihiro NakamuraAtsusuke YatomiYusuke SonodaHiroyuki TakaharaKyoko YamamotoYuya SuzukiKenichi TaniHidehiro IwaiYusuke NakanishiKen-ichi Hirata
Author information
JOURNAL OPEN ACCESS FULL-TEXT HTML Advance online publication
Supplementary material

Article ID: CJ-22-0065

Details
Abstract

Background: Several algorithms have been proposed for differentiating the right and left outflow tracts (RVOT/LVOT) arrhythmia origins from 12-lead electrocardiograms (ECGs); however, the procedure is complicated. A deep learning (DL) model, a form of artificial intelligence, can directly use ECGs and depict the importance of the leads and waveforms. This study aimed to create a visualized DL model that could classify arrhythmia origins more accurately.

Methods and Results: This study enrolled 80 patients who underwent catheter ablation. A convolutional neural network-based model that could classify arrhythmia origins with 12-lead ECGs and visualize the leads that contributed to the diagnosis using a gradient-weighted class activation mapping method was developed. The average prediction results of the origins by the DL model were 89.4% (88.2–90.6) for accuracy and 95.2% (94.3–96.2) for recall, which were significantly better than when a conventional algorithm is used. The ratio of the contribution to the prediction differed between RVOT and LVOT origins. Although leads V1 to V3 and the limb leads had a focused balance in the LVOT group, the contribution ratio of leads aVR, aVL, and aVF was higher in the RVOT group.

Conclusions: This study diagnosed the arrhythmia origins more accurately than the conventional algorithm, and clarified which part of the 12-lead waveforms contributed to the diagnosis. The visualized DL model was convincing and may play a role in understanding the pathogenesis of arrhythmias.

Idiopathic outflow tract ventricular arrhythmias (OT-VAs) include ventricular premature contractions (VPCs) and ventricular tachycardias (VTs) that usually originate from the right or left outflow tracts (RVOT/LVOT).1,2 Radiofrequency catheter ablation (RFCA) is an effective treatment option in patients with symptoms or a worsened cardiac function due to frequent OT-VAs.36 Predicting the anatomical origin of OT-VAs before the RFCA is important because it will help in guiding the access for the ablation, avoiding some complications, and saving the procedure and fluoroscopic time.7

Editorial p ????

Several algorithms have been proposed for differentiating RVOT or LVOT origins using 12-lead electrocardiograms (ECGs).820 Although those algorithms can classify the origins with a high accuracy, the procedure is time-consuming because it requires an evaluation of the QRS waveform in several leads using complicated algorithms.21 Recently, machine learning models that use waveform features as the input have been reported, but they are complicated because they require conversion processes from the waveforms to feature variables.

In recent years, the application of artificial intelligence (AI) in medical practice has grown dramatically. Deep learning (DL) models, which are forms of AI, have been applied to various aspects of medicine, and in the cardiovascular field, the usefulness of ECG in diagnosing ventricular dysfunction, valvular disease, and cardiomyopathy has been reported.22,23 However, DL models may not explain the evidence for the diagnosis because a DL model is considered a black-box method.24 To solve those problems, a DL model that could directly use the ECG morphology without any conversion process and could depict the importance of the leads and waveforms would be very useful. The purpose of this study was to create a DL model that could classify OT-VA origins more easily and better than conventional algorithms would, and to visualize its diagnostic contribution using the gradient-weighted class activation mapping (Grad-CAM) method. This method uses the gradient of a final convolutional layer to generate a coarse localization map that highlights the important regions in the image or wave.25

Methods

Study Population

A total of 80 cases, including 15 with OT-VAs originating from the LVOT and 65 from the RVOT, were collected from January 2009 to September 2021. We collected the preoperative 12-lead ECG data from the patients who underwent ablation therapy for OT-VAs. Patients whose OT-VA morphology could not be observed during the ablation procedure were excluded from this study. All methods were carried out in accordance with the relevant directives and regulations, as well as the Declaration of Helsinki, and informed consent was obtained from all the participants in the experiments. This clinical study was approved by the ethical review board of the Kobe University Medical Ethical Committee (No. B210168) on September 27, 2021.

Mapping and Ablation Protocol

We performed the ablation procedure with patients who were not sedated. Electroanatomic mapping systems such as the CARTO3 (Biosense Webster, Diamond Bar, CA, USA) or Ensite (Abbott, ST. Paul, MN, USA) were used. Intracardiac echocardiography assisted in defining the anatomy, facilitating the mapping, and assessing the catheter contact.

A 6-F quadripolar catheter was inserted via the femoral vein and placed through the atrioventricular valve to map the largest His potential. A standard 10-pole diagnostic catheter was positioned in the coronary sinus, and a 2 Fr catheter was inserted into the anterior interventricular vein (AIV). Pace and activation mapping were performed with a 7-French, 4-mm-tip non-irrigated or 7.5-French, 3.5-mm-tip irrigated ablation catheter in both the right and left ventricles to locate the origin of the OT-VA. When few OT-VAs were observed at the beginning of the electrophysiologic study, induction of the OT-VA was attempted by burst pacing from the RVOT or right ventricular apex with or without an isoproterenol infusion.

Activation mapping was performed in all patients during the OT-VAs. Pace mapping was performed at the maximum output (1.0 ms, 20 V), and the output was decreased until the pacing could not capture the myocardium. The ablation site was determined by matching the pace mapping (>11/12 leads) with the earliest bipolar ventricular electrogram preceding the QRS onset, the initial QS morphology for unipolar ventricular electrogram during OT-VAs. If a suitable ablation site in the RVOT was not located or ablation failed to abolish the arrhythmia, further mapping in the LVOT was performed via a retrograde aortic approach. The radiofrequency current was delivered with an ablation catheter, with a power setting of 25–40 W and a temperature limit of 43℃. Contact force-sensing catheters were used and the operators targeted a 5–30 g contact force. If the OT-VAs disappeared or the frequency of the arrhythmias diminished after the first 30 s of ablation, the energy was delivered continuously for 60 s. Ablation success was defined as the absence of spontaneous or induced OT-VAs at 30 min after the last energy delivery, and was confirmed by continuous cardiac telemetry over the subsequent 24 h of inpatient care. The origin of OT-VAs was determined based on the prematurity from QRS onset and the response to ablation. Systemic anticoagulation was achieved with intravenous heparin targeting a minimum activation clotting time of 350 s.

Data Preprocessing

In this study, the data set consisted of OT-VA ECG morphologies from 80 patients, and the origins of the OT-VAs were detected by electrophysiological tests. The 12-lead ECG data used in this study was a body surface ECG, with a sampling rate of 500 Hz and potential unit of µv, as in the general ECG format. Of the ECG data acquired from the ECG device in the Medical Waveform Format Encoding Rule (MFER) format, 800 ms including 1 heartbeat of the OT-VA was cropped. MFER is a widely used International Standards Organization standard for medical waveform data such as ECG, electroencephalogram, and blood pressure waveforms. Even if fragments of the morphologies before and after were left in the clipped data, they were left as they were. For all cases, the waveform of a single heartbeat of the OT-VA was cut out from the 12 leads and converted into a matrix of 400×12 data points per sample. A diagnosis was then made to determine whether the right OT or left OT was converted into binary variables.

Development of the DL Model

In this study, we used the 1-dimensional convolutional neural network (CNN) model, often used as a waveform recognition model. The overview of the DL model is shown in Figure 1. The training and test datasets were split by using a 5-fold cross-validation method. Therefore, 80% of the entire dataset was used as the input data. The input shape was (N, 400, 12) and the output shape was (N, 2), where N stands for the number of batch sizes. There were 11 layers in the entire network model, including the CNN layer with an activation function and pooling layer, and the final layer was a sigmoid function. Adam was used as the optimizer, and the default values were used for the parameters. Bayesian optimization was used to optimize the hyperparameter. All the models were implemented with the Tensorflow framework.

Figure 1.

Overview of the deep learning architecture. The matrix of the 12-lead ECGs is used as the input for the deep learning model to classify the binaries. The gradient-weighted class activation mapping (Grad-CAM) method is used to calculate the contribution rate and heat maps of each induction based on the gradient of the weights of the deep learning model. 1D-CNN, 1-dimensional convolutional neural network; ECG, electrocardiogram; LVOT, left ventricular outflow tract; RVOT, right ventricular outflow tract.

Explainable DL Model

In this study, we used the Grad-CAM method to visualize the model, which could output a heat map of which parts of the data contributed to the results using the gradient of the last activation layer. Generally, the Grad-CAM is used for 2-dimensional images; however, we modified and applied the method to the 1-dimensional data in this study.

To compute the contribution ratio of each lead, the maximum values of each lead were normalized. The contribution ratio, that is, which part of the leads highly contributed to the diagnosis, was calculated for each sample.

Conventional Algorithm

Several algorithms have been reported for diagnosing the origin of arrhythmias from the 12-lead ECG. To compare the performance of our model implemented in this study, we used a conventional algorithm that was often used clinically.13,26 This algorithm was based on the precordial transition pattern seen during clinical arrhythmias vs. sinus rhythm to differentiate VAs arising from the RVOT vs. those originating from the LVOT. The R- and S-wave amplitudes in lead V2 were measured during both sinus rhythm and the VAs. The transition ratio was then calculated by computing the percentage R-wave during the VT (R / R + S)VT divided by the percentage R wave in sinus rhythm (R / R+ S)SR. If the transition ratio was <0.6, then a RVOT origin was likely. If the transition ratio was ≥0.6, then the LVOT origin was likely.

In the comparison between the conventional method and AI model, the metrics of each model were examined for any statistically significant differences using a 1-sample t-test.

Results

Patient Characteristics and Procedural Data

During the study period, 80 patients who underwent RFCA of OT-VAs were studied. The baseline patient characteristics are summarized in Table 1. The mean age was 49±16 years, and 36 patients (45%) were males. The average left ventricular ejection fraction was 59±9%. The VA origin was in the RVOT in 65 patients (81.2%) and LVOT in 15 (18.8%). The details of the different origins are shown in Supplementary Table. Among the patient characteristics and procedural data, the left ventricular end-diastolic diameter did not differ between the RVOT and LVOT groups (LVDd: 48±6 vs. 52±5, P=0.01).

Table 1. Baseline Patient Characteristics
  All
(n=80)
RVOT
(n=65)
LVOT
(n=15)
P value
Age, years 49±16 48±16 50±15 0.62
Male 36 (45) 27 (42) 9 (60) 0.20
ICD 1 (1) 1 (2) 0 (0) 0.63
VT 10 (12) 9 (14) 1 (7) 0.45
Syncope 10 (12) 10 (15) 0 (0) 0.11
LVEF 59±9 60±9 57±9 0.33
LVDd 49±6 48±6 52±5 0.01
THB, n×103/24 h 109.5±20.9 109.5±21.9 109.8±16.7 0.95
VPC, n×103/24 h 25.9±1.7 24.6±1.7 31.2±1.6 0.16
β-blocker 52 (65) 27 (42) 9 (60) 0.46
Antiarrhythmic drug 4 (5) 4 (6) 0 (0) 0.33
Procedure time, min 175±50 171±52 189±43 0.17
Total radiofrequency lesions delivered 15±9 16±9 12±9 0.11

Data are presented as the mean±SD or n (%). ICD, implantable cardioverter defibrillators; LVDd, left ventricular end-diastolic diameter; LVEF, left ventricular ejection fraction; LVOT, left ventricular outflow tract; RVOT, right ventricular outflow tract; THB, total heart beats; VPC, ventricular premature contraction; VT, ventricular tachycardia.

Comparison of the Arrhythmia Origin Prediction Between the DL Model and Conventional Algorithm

The results of an average of 10 OT-VA origin predictions by the DL model and the predictions by the conventional algorithm are shown in Table 2. The average prediction of the DL model was 89.4% (88.2–90.6) for accuracy, 95.2% (94.3–96.2) for recall, and 93.6% (92.9–94.3) for the F1-Score, which were significantly better than that of the conventional algorithm. In contrast, the precision of the conventional algorithm was better than that of the DL model.

Table 2. Comparison of Each Metric Between the Conventional Model and DL Model
  Accuracy Precision Recall F1-Score
Average of the DL model 89.4% (88.2–90.6) 92.0% (91.0–93.0) 95.2% (94.3–96.2) 93.6% (92.9–94.3)
Conventional algorithm 71.3% 95.7% 67.7% 79.3%
P value <0.01 <0.01 <0.01 <0.01

Each number represents a metric (upper 95% confidence interval–lower 95% confidence interval). F1-Score = 2 × Precision × Recall / (Precision + Recall). DL, deep learning.

Focus on the 12-Lead Morphology for the Arrhythmia Origin Prediction of the DL Model

Figure 2 shows the proportion of leads contributing to the diagnosis predicted by the DL model. The proportions of contribution to the prediction were higher than the average for the limb leads and leads V1 and V2 for the RVOT origins and limb leads and leads V1, V2, and V3 for the LVOT origin. Moreover, the distribution of the ratios differed between the left and right origins. Although leads V1 to V3 and limb leads had a balanced focused in the LVOT group, the contribution ratio in the limb leads, especially leads aVR, aVL, and aVF, was higher in the RVOT group. Comparing the Q-wave amplitude in leads aVR and aVL, and the R-wave amplitude in lead aVF, the Q-wave amplitude in leads aVR and aVL did not differ between the 2 groups; however, the R-wave amplitude in lead aVF was significantly higher in the LVOT origin (Table 3).

Figure 2.

This graph shows the proportion of the contribution to the deep learning (DL) model. The results showed that the contribution of the regions for the prediction differed between the RVOT and LVOT origins: leads V1 to lead V3, and limb leads are equally contributed in the LVOT group, whereas the limb leads, especially leads aVR, aVL, and aVF, were emphasized more in the RVOT group. The error bar is expressed as the standard deviation. LVOT, left ventricular outflow tract; RVOT, right ventricular outflow tract.

Table 3. Q and R-Wave Amplitude Comparison
  All
(n=80)
RVOT
(n=65)
LVOT
(n=15)
P value
Q-wave amplitude in aVR (mV) 8.9±2.6 8.8±2.6 9.4±2.7 0.43
Q-wave amplitude in aVL (mV) 8.5±3.8 8.1±3.9 10.2±3.6 0.05
R-wave amplitude in aVF (mV) 17.1±5.0 16.3±4.7 20.5±6.2 <0.01

Data are presented as mean±SD. Abbreviations as in Table 1.

A representative case is shown in Figure 3. This patient underwent a successful ablation in the RVOT (within the pulmonary artery). Although the conventional algorithm predicted the origin as being in the LVOT, the DL model predicted the origin was in the RVOT (prediction value: 91.3%). We created a heat map to visualize which part of the morphology was focused on (Figure 3). The upstroke of the QRS complex was highlighted in the precordial leads, and the entire QRS complex was highlighted in the limb lead, especially leads V1 to V4.

Figure 3.

(A) Shown are the 12-lead waveforms during sinus rhythm and ventricular premature contractions in a representative patient (Transition ratio = C / [C + D] ÷ A / [A + B] = 0.8 / [0.8 + 1.9] ÷ 0.7 / [0.7 + 1.1] = 0.76 [≥0.6]). (B) A heat map of the parts of each waveform that were focused on in the deep learning model. The darker the blue, the more it contributed to the diagnosis. This case was predicted to have an RVOT origin with a probability of 91.3%. RVOT, right ventricular outflow tract.

Discussion

In this study, we developed a DL model that could more easily and better classify the origins of OT-VAs using the ECG morphologies. In addition, we visualized the predictions of the DL model by highlighting the regions that contributed to the prediction using the Grad-CAM method. The results showed that the contribution of the regions for the prediction differed between the RVOT and LVOT origins: leads V1 to lead V3 and limb leads equally contributed in the LVOT group, whereas the limb leads, especially leads aVR, aVL, and aVF, were more emphasized in the RVOT group.

Comparison of the Prediction Between the DL Model and Conventional Algorithms

There are many algorithms for predicting the origin of arrhythmias, and some conventional algorithms require evaluating multiple leads.1520 Complexity is necessary to increase the accuracy, but the complexity makes it difficult to use in clinical practice. For the comparison of the DL model, we used an algorithm that required only a single lead, was relatively easy to handle, and is frequently used in clinical practice.13,26 The representative case is presented in Figure 3 and originated within the pulmonary artery (PA) in the RVOT. One of the reasons why the conventional algorithm was incorrect in its prediction in this case is that the conventional algorithm may not have been able to provide evidence to support the PA origin.27 Conventional algorithms sometimes require the use of a combination of algorithms specific for predicting the origin. Some previous reports have used machine learning for an origin prediction, but they are similar to the conventional algorithms in terms of the complexity, because they require manual extraction of the morphological features.21 In contrast, the DL model can be applied easily in clinical practice because the ECG morphology can be an input. It is used in various areas, such as in accessory pathway analyses in patients with Wolff–Parkinson–White syndrome.28

Although the precision of the conventional algorithm was higher than that of the DL model, all the other metrics of the DL model were significantly higher than that of the conventional algorithm. That may have been because the DL model uses the full lead information and also has enough expressive capacity to handle high-dimensional data.

Implication of Leads, aVR and aVL, for the Arrhythmia Origin Prediction

In this study, none of the leads had a remarkably high contribution, which suggests that the diagnosis with AI requires information from many leads. However, the distribution of the lead contribution, especially of leads aVL and aVR, differed between the RVOT and LVOT origins in the DL model prediction. As the arrhythmia origin prediction is reportedly based on the location of the transition zone, precordial leads have been mainly used in conventional algorithms. However, the visualized DL model showed that leads aVR, aVL, and aVF were also important, especially in diagnosing RVOT origins. Kamakura et al reported that when the QS wave depth in lead aVL was larger than that in lead aVR, the origin was likely to be in the LVOT, and when the QS amplitude in lead aVR was equal to or larger than that in lead aVL, the origin was likely to be in the RVOT.8 Applying this algorithm to our patient group, the accuracy for the entire group, LVOT group, and RVOT group was 63.8%, 57.9% (22/38), and 92.8% (39/42), respectively, suggesting that leads aVL and aVR contributed to the prediction of the RVOT origins.

The ECG morphology in leads aVR and aVL were similar to those recorded from the unipolar ECGs in the right and left shoulder areas, respectively. As Figure 4 shows, in the VAs from the right side of the RVOT, lead aVR exhibited a larger QS wave than lead aVL. In contrast, in the VAs from the left side of the RVOT and LVOT, lead aVL exhibited a larger QS wave than lead aVR because the VAs moved away from the left upper region of the heart. We believed that those mechanisms were why aVR and aVL were also important in predicting the VA origins.

Figure 4.

In the VAs from the right side of the RVOT (yellow circle), lead aVR had larger QS waves than lead aVL. In contrast, in the VAs from the left side of the RVOT and LVOT (yellow dot circle), lead aVL had larger QS waves than lead aVR because the VAs moved away from the left upper region of the heart. VA, ventricular arrhythmias.

Implications of Lead aVF in the Arrhythmia Origin Prediction

It is less common to use lead aVF to distinguish between the RVOT and LVOT. The limb bipolar leads are measured as a potential between 2 points on the limb. Lead aVF is expressed as (2 × lead II − lead I) / 2 using only leads I and II; that is, the height of the VA origin and axis of the conduction direction.29 Lead I tends to be positive with RVOT origins and negative with LVOT origins,8 and the wave height of lead aVF tends to be lower with RVOT origins and higher with LVOT origins. In the present study patient group, the R-wave amplitude in aVF was significantly higher for LVOT origins than RVOT origins (Table 3), suggesting that lead aVF may also be useful for differentiating between the 2 groups.

Study Limitations

In this study, the number of patients was low, therefore we could only classify the arrhythmia origin into 2 classes. Also, this study did not contain a prospective validation cohort. Future studies need to increase the number of cases from multiple institutions and number of classifications.

The RVOT and LVOT origins were determined by the difference in the prematurity of the intracardiac electrograms in the electroanatomic mapping. In reality, the RVOT and LVOT origins may be deeper and influenced by preferential pathways.

Conclusions

We can now diagnose the origin of VAs more accurately than by using a conventional algorithm and clarify which part of the 12-lead waveform contributed to the diagnosis using the Grad-CAM method. The visualized DL model is convincing and may play a role in understanding the pathogenesis VAs.

Acknowledgment

The authors would like to thank Mr. John Martin for his linguistic assistance.

Author Contributions

K.N.: clinical practice/data sampling/drafting article; M.N.: concept/design/data analysis/drafting article; K.K.: concept/design/data analysis/interpretation/drafting article; M.S.: collection/statistics/drafting article; K.F.: data collection/statistics; M.T.: data collection/statistics; M.E.H.: data collection/statistics; P.S.: data collection/statistics; J.S.: data collection/statistics; T.N.: data collection/statistics; A.Y.: data collection/statistics; Y. Sonoda: data collection/statistics; H.T.: data collection/statistics; K.Y.: data collection/statistics; Y. Suzuki: data collection/statistics; K.T.: data collection/statistics; H.I.: data collection/statistics; Y.N.: data collection/statistics; K.H.: approval of the article for publication.

Disclosures

The Section of Arrhythmia is supported by an endowment from Abbott JAPAN and Medtronic JAPAN and has received a scholarship fund from Biotronik JAPAN. K.H. chairs the Section and K.F., K.K. belong to the Section. However, all authors report no conflict of interests for this manuscript’s contents.

K.H. is a member of Circulation Journal’s Editorial Team.

Sources of Funding

There are no sources of funding related to this study.

Data Availability

The data, analytic methods, and study materials will not be made available to other researchers for purposes of reproducing the results or replicating the procedure.

IRB Information

Kobe University Medical Ethical Committee (Reference number: No. B210168) approved this study.

Supplementary Files

Please find supplementary file(s);

http://dx.doi.org/10.1253/circj.CJ-22-0065

References
 
© 2022, THE JAPANESE CIRCULATION SOCIETY

This article is licensed under a Creative Commons [Attribution-NonCommercial-NoDerivatives 4.0 International] license.
https://creativecommons.org/licenses/by-nc-nd/4.0/
feedback
Top