Introduction

Premature ventricular complexes (PVCs) are common arrhythmic events associated with an increased risk of reduced left ventricular (LV) ejection fraction, heart failure, and all-cause mortality.¹ In patients with structurally normal hearts, idiopathic PVCs most frequently originate from the right or left ventricular outflow tracts (RVOT/LVOT). When pharmacologic therapy is ineffective or poorly tolerated, catheter ablation emerges as an effective and well-established treatment option. Outflow tract ventricular arrhythmias can be safely and efficiently targeted using radiofrequency (RF) catheter ablation, which achieves high long-term success rates.² Accurate focus localization is essential, yet RVOT–LVOT differentiation remains difficult due to their close anatomical proximity—especially near the aortomitral continuity—resulting in limited accuracy of current electrocardiographic (ECG)-based algorithms. Although several ECG algorithms have been developed to aid in PVC localization, many are based on fixed cutoff values derived from small cohorts, and their applicability across diverse populations remains inconsistent.^3-9 Existing ECG algorithms may be insufficient in complex cases, prompting us to re-evaluate key features and develop an integrated logistic regression-based supervised learning approach model. Supervised learning approach is a method that learns patterns from labeled data, much like a clinician recognizing risk by reviewing past cases with known outcomes. Using this approach, validated ECG markers were combined into a simple and interpretable tool for distinguishing RVOT from LVOT PVCs. A supervised learning approach model was developed based on logistic regression to distinguish RVOT from LVOT PVCs, and it was compared with 7 prior algorithms, aiming to provide a more accurate and practical tool for PVC localization and ablation planning.

Methods

Dataset

The study retrospectively reviewed 370 patients who underwent PVC ablation between January 2015 and June 2020. Exclusion criteria included the absence of electrophysiologic study data, structural heart disease, baseline bundle branch block, non-outflow tract PVCs, unsuccessful ablation, and ECGs with evidence of improper V1–V2 lead placement. Appropriate precordial lead positioning was confirmed using indirect ECG criteria, including physiologic P-wave morphology, normal precordial R-wave progression, and the absence of patterns suggestive of high chest placement (e.g., unusually dominant R waves or atypical rSr′ in V1–V2). A total of 22 ECGs were excluded due to suspected improper V1–V2 lead placement, based on the predefined ECG criteria. After applying these criteria, 116 patients with LBBB morphology and an inferior axis were included in the final dataset for model development (Figure 1). To ensure methodological consistency, the study included only PVCs with an LBBB morphology and inferior axis, which is the phenotype used in most previously published RVOT–LVOT differentiation algorithms. This approach allowed direct comparison with existing tools and prevented heterogeneity related to other PVC morphologies.

This study was conducted in accordance with institutional and national ethical standards. Approval for the retrospective analysis was obtained from the Institutional Ethics Committee on January 19, 2021 (Approval Number: GO 21/79, 2021/02-43). As the study involved a retrospective review of clinical records acquired during routine care, the requirement for additional study-specific patient consent was waived. All patients had previously provided informed consent for their ablation procedures, and no directly identifiable personal information was used during data analysis.

Electrophysiologic Study and Ablation Procedure

All antiarrhythmic pharmacotherapies were terminated at least 5 half-lives before the ablation to guarantee drug elimination. The ablation procedure was performed only under local anesthesia. In all patients, a coronary sinus catheter was placed. Electroanatomical mapping was performed using a 3.5-mm open-irrigated ThermoCool catheter (Biosense Webster, Diamond Bar, CA, USA) in conjunction with the CARTO 3 (Biosense Webster, Irvine, CA, USA) or EnSite Precision (Abbott Medical, St. Paul, MN, USA) 3D mapping systems. The earliest activation site was determined from the surface QRS, and precocity was manually confirmed using the 3D mapping system and the Prucka/CardioLab EP recording system (GE Healthcare, Milwaukee, WI, USA). Radiofrequency energy was delivered to the earliest activation site, and the region where PVC originated was identified by the termination of PVC with RF energy. Additionally, heparin was given intravenously to keep the active clotting time over 300 seconds during mapping in the LVOT. Acute ablation success was defined as the complete elimination of the targeted clinical PVC during the electrophysiology procedure and persisting for at least 30 minutes after the final RF energy application, both at baseline and following isoproterenol infusion. Additionally, the absence of PVCs was confirmed through continuous full-disclosure telemetry monitoring for 24 hours post ablation during inpatient observation. Long-term procedural success was defined as a ≥80% reduction in the burden of the targeted PVC morphology on 24-hour Holter monitoring at 3-6 months of follow-up, in the absence of antiarrhythmic drug escalation or repeat ablation. All patients were routinely followed with outpatient clinic visits, including ECG and HF evaluation at 1, 3, and 6 months. Any recurrence of clinical PVCs was also assessed symptomatically and via rhythm monitoring, as appropriate.

Electrocardiographic Analysis

All ECG measurements and algorithm calculations were performed manually by a single experienced observer who was blinded to the ablation results, following the exact definitions and cutoff criteria provided in the original publications. A standard sweep speed of 25 mm/s and a voltage calibration of 10 mm/mV 12-lead ECGs were obtained for all patients during sinus rhythm and episodes of PVCs (Figure 2). These recordings were conducted using the conventional arrangement of the limb and chest leads. Special attention was given to the accurate placement of the electrodes for leads V1 and V2 in the fourth intercostal space to avoid erroneous positioning, which could significantly alter the QRS morphology associated with outflow tract ventricular arrhythmias. LBBB morphology was defined in accordance with American Heart Association / American College of Cardiology Foundation / Heart Rhythm Society recommendations, including a predominantly negative QRS complex in lead V1 (S wave amplitude exceeding the R wave) and a broad, notched, or slurred R wave in lateral leads.¹⁰ An inferior axis was defined by a positive net QRS vector in leads II, III, and aVF, consistent with standard ECG criteria. Measurements were conducted using an electronic caliper within the recording system. To differentiate between right and left ventricular origins of PVCs, several ECG algorithms have been published. In the study, an analysis was conducted using 7 algorithms that are the most frequently used in previous studies, aiming to evaluate their effectiveness and accuracy in localizing the origin of PVCs. A total of 7 previously published algorithms were selected based on their availability at the time of study design (2020–2022), their validation in LBBB–inferior axis outflow-tract PVC cohorts, and their reliance on standard 12-lead ECG measurements. Features extracted from those 7 algorithms were used for model input. The selected parameters included:

Statistical Analysis

Statistical analyses were conducted using IBM SPSS Statistics for Windows, version 26.0 (IBM Corp., Armonk, NY, USA) with a 95% CI. For numerical variables, mean (M), standard deviation (SD), minimum, maximum, and median values are reported; for categorical variables, frequencies (n) and percentages (%) are presented. The chi-square test was used to evaluate relationships between categorical variables. Between-group comparisons were performed using independent i-tests or Mann–Whitney U-tests, based on the distribution of the data. Normality was assessed using both skewness/kurtosis values (with thresholds between ±3) and formal tests including the Shapiro–Wilk and Kolmogorov–Smirnov tests. Parametric or non-parametric methods were applied accordingly.

To identify key predictors of PVC origin and develop a classification model, a supervised learning approach approach based on binary logistic regression was employed. The classification problem was defined by labeling the left LVOT origin as 1 and the RVOT origin as 0. All candidate predictor variables—derived from established ECG algorithms—were initially included in the model. Because LVOT and RVOT cases were relatively balanced in the dataset, no class-balancing techniques or resampling were applied, and the full dataset was used for model derivation without a train–test split or cross-validation.

To optimize model performance and reduce overfitting risk, a backward stepwise logistic regression analysis (Wald test-based) was used. Through 8t iterative steps, non-significant variables were systematically removed based on predefined statistical criteria (P > .05). The final model retained only those predictors with statistically significant contributions, and model fit at each step was evaluated using the likelihood-ratio chi-square test. The statistical significance of the final model (P < .05) confirmed its robustness and predictive validity.

The diagnostic performance of the newly derived machine learning model was assessed using receiver operating characteristic (ROC) curve analysis. In addition to the new model, ROC analyses were performed for 7 previously published ECG-based algorithms. The area under the curve (AUC) was calculated for each method, and pairwise AUC comparisons were conducted using the DeLong test. To evaluate the overall discriminative power of each model, the Youden Index was calculated. These performance metrics were computed for both the entire patient cohort and the V3 precordial transition subgroup.

Results

The study included 116 participants, with 59 (51%) males. The average age was 45.8 years (±14.6). Most participants (87.7%) had no history of ablation, while 12.3% had undergone the procedure before. Of the 116 patients in the study, 63 (54.3%) underwent ablation for LVOT-origin PVCs, while 53 (45.7%) had RVOT-origin PVCs (Table 1).

Among the 63 patients who underwent LVOT ablation, the most targeted site was the left coronary cusp (LCC) in 21 patients (33.3%), followed by the right-left coronary cusp (RCC-LCC) junction in 17 patients (27.0%), and the LV summit region in 13 patients (20.6%). In the RVOT ablation group (n = 53), the most frequently ablated site was the RVOT anteroseptal region in 15 patients (28.3%), followed by the RVOT posteroseptal region in 10 patients (18.9%). Additionally, the RVOT anterior and posterior regions were each targeted in 9 patients (17.0%). This distribution illustrates the predominant ablation sites within the LVOT and RVOT groups among the study cohort. In the population, 13 patients (20.6% of LVOT cases) were classified as having PVCs originating from the LV summit. However, none of these ablations were performed via an epicardial approach. Instead, successful ablation was achieved from adjacent endocardial locations such as the aortic cusps or the subvalvular LVOT, where earliest activation was recorded and effective RF energy delivery was possible. These areas likely represent endocardial breakout points of summit-originating PVCs.

Quantitative features derived from algorithms that discriminate right or left origin for PVC were compared between RVOT and LVOT ablation groups (Table 2). The RVOT group demonstrated statistically higher values in PVC TZ score, PVC QRS duration, PVC S wave amplitude in V1, PVC QRS amplitude in V1, PVC S wave amplitude in V2, and PVC QRS amplitude in V2, indicating greater wave amplitude and QRS complex duration in this group. Conversely, the LVOT group exhibited significantly higher values in the PVC R wave deflection interval in V3, PVC R wave amplitude in V2, and PVC R wave amplitude in V3, reflecting distinct patterns of R wave morphology in this region (Table 3).

The diagnostic performance of feature-based classification models was assessed using standard evaluation metrics, including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). These metrics were calculated both for the overall dataset and for a clinically relevant subgroup of patients with a precordial transition at lead V3. In the full cohort, the Combined TZ Index–based model achieved the highest sensitivity (90.5%) and NPV (85.4%), indicating a robust capacity for identifying true LVOT cases while minimizing false negatives. Alternatively, the model incorporating the R Wave Duration Index and R/S Amplitude Index yielded the highest specificity (98.1%) and PPV (94.7%), reflecting strong discriminatory power for RVOT origin and low false positive rates. When applied to the V3 transition subgroup, the Combined TZ Index again demonstrated superior sensitivity (95.2%) and NPV (82.4%), whereas the R Wave Duration Index and R/S Amplitude Index model achieved perfect specificity and PPV (100% each), effectively eliminating false positives within this subgroup (Table 4).

All ECG algorithms showed statistically significant effectiveness in predicting the exact ablation location, confirming their utility in clinical settings (Table 5).

Building upon the diagnostic insights derived from individual ECG algorithms, a supervised learning approach model was developed using logistic regression to integrate the most informative features. The final model incorporated 4 key predictors: PVC TZ score, PVC V2 R wave duration, PVC V3 R wave amplitude, and PVC V2 S wave amplitude—all of which demonstrated independent discriminatory power in differentiating LVOT from RVOT origins (Table 6).

The logistic regression function was defined as: Y = 8.436 − 2.036 × PVC TZ score − 0.06 × PVC V2 R wave duration + 4.661 × PVC V3 R wave amplitude − 1.958 × PVC V2 S wave amplitude, with a threshold of Y > 0.5 indicating an LVOT origin (Figure 3; Table 7).

Model validation showed that 84.4% of LVOT cases and 82.7% of RVOT cases were accurately classified, yielding a statistically significant distinction between groups (P < .05). The model achieved a sensitivity of 84.4%, specificity of 82.7%, PPV of 85.7%, and NPV of 81.1% (Table 8). Moreover, the Youden Index (0.66) exceeded that of the widely used V2S/V3R index. In the V3 transition subgroup, the model maintained a sensitivity of 92.1% and an NPV of 81.1%, with a Youden Index of 0.46—matching the V2S/V3R index and supporting the robustness of the model across challenging ECG patterns (Table 4).

Receiver-operating characteristic curve analyses were performed to evaluate and compare the classification performance of the 7 previously published algorithms and the newly developed supervised learning approach model. The new model demonstrated an AUC of 0.942 (Figure 4), significantly outperforming the AUCs of all individual algorithms in the overall population (Table 9). In the V3 transition subgroup, the model maintained a high discriminative ability with an AUC of 0.878 (Figure 5), again showing statistically superior performance compared to the pre-existing algorithms (Table 10). These findings further support the robustness and generalizability of the proposed machine learning–based approach across both general and anatomically challenging subgroups.

Discussion

In this study, the diagnostic performance of 7 previously published ECG-based algorithms were systematically compared with a newly developed supervised learning approach model using logistic regression. The new model integrated 4 significant predictors: PVC TZ score, PVC V2 R wave duration, PVC V3 R wave amplitude, and PVC V2 S wave amplitude.

Based on ROC curve analysis, the new algorithm achieved the highest discriminative ability with an AUC of 0.942 in the overall cohort and 0.878 in the V3 precordial transition subgroup. These values were significantly superior to all comparator algorithms (P < .001 for all), including the widely used V2S/V3R index (AUC = 0.103 and 0.158, respectively), the combined TZ and V2S/V3R index (AUC = 0.852 and 0.716), and the R wave–based indices by Okuyan¹¹ (AUC = 0.630 and 0.575). These results underscore the model’s robustness across both general and anatomically challenging patient populations.

Collectively, these findings suggest that the integration of selected ECG features through a supervised learning model not only enhances classification accuracy across the full population but also sustains diagnostic performance in more nuanced ECG patterns such as precordial transition in V3. By contrast, rule-based algorithms, while useful, appear more limited in adaptability and may underperform when applied outside their original validation context.

Compared to traditional rule-based diagnostic methods, the supervised learning approach employed in this study offers several methodological and practical advantages.¹¹ By simultaneously evaluating and weighting multiple predictors based on statistical learning, this approach minimizes human bias in cutoff selection and improves the model’s generalizability across varied populations. Logistic regression, a widely used supervised classifier, provides both predictive power and interpretability—2 qualities critical in clinical decision-making.¹² The model’s consistent performance across the general cohort and the V3 precordial transition subgroup illustrates how data-driven approaches can enhance precision and reliability in electrophysiologic diagnostics. These findings are in line with previous reports demonstrating the superiority of supervised learning over conventional heuristics in clinical prediction tasks.

The supervised learning approach model’s predictors were derived from PVC complexes, not sinus beats, proving ectopic ventricular activity alone can accurately differentiate between LVOT and RVOT origins. This emphasizes the importance of PVC-specific characteristics for data-driven localization tools. The model also does not rely on sinus rhythm–based precordial transition, which can be affected by cardiac rotation.¹³ By focusing only on PVC features, the model is less impacted by anatomical variations and may be more reliable in diverse patient anatomies.

The present study reinforces the growing body of evidence suggesting that supervised learning approaches can outperform traditional rule-based algorithms in electrophysiological classification tasks. While rule-based models often rely on empirically determined cut-off values that may not generalize well across diverse patient cohorts, supervised models optimize parameter weighting based on statistical learning from labeled data.¹⁴ This allows for improved adaptability to anatomical variability, particularly in anatomically ambiguous regions such as the aortomitral continuity.

An additional strength of the current model lies in its use of logistic regression, which provides not only robust predictive performance but also interpretability, an essential attribute in clinical decision-making. Unlike complex “black-box” models such as neural networks or ensemble classifiers, logistic regression enables clinicians to understand how each ECG variable contributes to the final classification.¹⁵ This interpretability facilitates model acceptance among electrophysiologists and supports potential integration into clinical decision support systems.

The model’s consistent performance across the general cohort and the V3 precordial transition subgroup illustrates how data-driven approaches can enhance precision and reliability in electrophysiologic diagnostics. A recent study by Zheng et al¹⁵ proposed a machine learning-based method for localizing the origin of outflow tract PVCs using 12-lead ECG data. While both their study and ours share a common goal of improving diagnostic precision through machine learning, key methodological differences exist. Zheng et al¹⁵ employed an ensemble learning model based on XGBoost, a decision tree–based algorithm, incorporating more than 50 automatically extracted ECG features. In contrast, the approach utilized a supervised learning model constructed with a smaller number of clinically established features derived from 7 previously published ECG algorithms. In addition, the recent systematic review by Moreno-Sánchez et al¹⁴ highlighted the rapid expansion of artificial intelligence (AI)-based ECG interpretation and emphasized the importance of interpretable, physiology-grounded models, further supporting the relevance of the feature-based supervised learning framework.

This distinction highlights a major strength of the method—clinical interpretability. Because each variable in the model corresponds to well-known diagnostic ECG criteria, the model’s decision-making process can be better understood and trusted by clinicians. Moreover, unlike the Zheng et al¹⁵ study, which did not assess model performance in clinically relevant subgroups, an additional analysis was conducted in patients with a transitional lead at V3, further demonstrating the robustness of the model across heterogeneous presentations. Lastly, the use of backward stepwise logistic regression based on the Wald statistic provided a statistically grounded feature selection process, resulting in a streamlined and explainable model architecture. In practical terms, the model can be applied by measuring 4 routinely available ECG parameters and entering them into the logistic regression formula to estimate LVOT probability. This could be easily implemented as a simple web-based calculator or integrated into ECG analysis or mapping systems for real-time use.

Recent advances in interpretable AI further underscore the transition from rule-based ECG criteria toward data-driven electrophysiologic analysis. For example, an ECG-only explainable deep-learning model developed by van de Leur et al¹⁶ successfully predicted malignant ventricular arrhythmias in phospholamban cardiomyopathy, demonstrating how contemporary AI systems can extract prognostic and morphological signals beyond traditional ECG heuristics. Similarly, Bhatia et al¹⁷ applied a convolutional deep-learning framework to intracardiac electrograms from subcutaneous Implantable Cardioverter Defibrillators (ICD) and showed that upstream intracardiac electrogram segments could predict ventricular arrhythmias with excellent performance, highlighting the potential of device-based signals for proactive, AI-driven arrhythmia prediction.¹⁷ These advances highlight the growing movement toward explainable and physiology-driven AI in cardiac electrophysiology. The study aligns with—and contributes to—this shift by demonstrating that interpretable, feature-based models can outperform traditional rule-based algorithms while remaining fully transparent and clinically usable.

One major limitation of this study is its retrospective and single-center design, which may restrict the external generalizability of the findings. In addition, the sample size (n = 116) is modest for a supervised learning analysis and may increase the risk of overfitting. Because the primary objective was to derive a classification model rather than to perform full model validation, the entire dataset was used for model development without a separate training–testing split or cross-validation. Accordingly, all reported AUC values were derived from the same dataset. Thus, the present work should be regarded as a derivation study rather than a fully validated predictive model. Similar to previously published ECG algorithms, this study represents a derivation model developed on a single dataset without cross-validation or an independent validation cohort.⁶ Consequently, the reported performance metrics should be interpreted with caution, and prospective validation in independent, multicenter cohorts is essential to assess the model’s robustness across diverse anatomical and demographic profiles. Furthermore, only cases with successful ablation were included to ensure definitive localization of PVC origin based on electrophysiologic and ablation endpoints. While this design strengthens internal validity, it may introduce selection bias and limit generalizability to patients who are not candidates for ablation or in whom ablation is unsuccessful.

Another limitation is the absence of interobserver variability analysis, as all ECG measurements were performed by a single investigator. While this approach ensured internal consistency, it precludes the assessment of reproducibility and operator dependence, which are critical factors for real-world application. Future studies should incorporate blinded dual-reviewer measurements with inter-rater agreement metrics such as Cohen’s kappa.

Although electroanatomical mapping and fluoroscopic guidance were employed to localize PVC origins during ablation, advanced imaging modalities such as intracardiac echocardiography (ICE) and cardiac magnetic resonance imaging (MRI) were not utilized. While ICE can enhance anatomical visualization and catheter-tissue contact in real time, particularly in complex ventricular anatomy, cardiac MRI may provide adjunctive information about myocardial fibrosis or intramural foci in selected patients. Although neither modality is routinely required for PVC ablation, their absence may have limited anatomical precision in certain challenging cases, such as those involving the LV summit or ambiguous epicardial sites.

It should also be noted that the diagnostic performance of previously published algorithms can vary significantly depending on the study population, operator experience, and methodological design. Therefore, even well-validated algorithms may perform differently across patient groups. Consistent with this, the findings of the study may not be directly generalizable to other populations without external validation.

Lastly, while the model developed in this study demonstrated superior diagnostic performance, it does not introduce entirely novel ECG parameters. Instead, it represents a re-analysis and re-integration of previously described ECG features using a supervised learning approach. This method allows for a more objective and data-driven synthesis of existing knowledge, enhancing diagnostic accuracy through statistical modeling rather than heuristic cutoff values.

Conclusion

The supervised learning approach model developed in this study—based on logistic regression—integrated 4 PVC-specific ECG features: TZ score, V2 R wave duration, V3 R wave amplitude, and V2 S wave amplitude. Compared to 7 previously established algorithms, the model demonstrated superior diagnostic accuracy, yielding the highest AUC in both the overall cohort and the diagnostically challenging V3 transition subgroup. Its data-driven framework, anchored in clinically interpretable variables, enables more precise and noninvasive localization of PVC origins, thereby enhancing the efficacy and targeting of catheter ablation procedures. Future work should focus on external validation across larger multicenter populations and the integration of this model into real-time clinical decision support systems.

Footnotes

Ethics Committee Approval: This study was approved by the Ethics Committee of Hacettepe University Faculty of Medicine on January 19, 2021 (Approval Number: GO 21/79, 2021/02-43).

Informed Consent: Informed consent was waived due to the retrospective design of the study.

Peer-review: Externally and internally peer-reviewed.

Author Contributions: Conception – A.S., C.Ç., A.K.; Design– A.S., C.Ç., A.K., U.C.; Supervision – H.Y., K.A., A.H.A.; Resources – H.Y., K.A.; Materials – N.O., A.K.; Data Collection and/or Processing– A.S., C.C., N.O.; Analysis and/or Interpretation – A.S., C.C., A.K., U.C.; Literature Review – A.S., C.Ç., N.O.; Writing – Original Draft: A.S., C.Ç., A.K.; Critical Review – All authors.

Acknowledgments: No generative artificial intelligence or AI-assisted technology was employed in the preparation, writing, or editing of this article. All content is the authors’ original work.

Declaration of Interests: The authors have no conflicts of interest to declare. Dr. Necla Ozer is a member of the Editorial Board of this journal. She was not involved in the review or decision process of this manuscript.

References

1. Keleş N, Kahraman E, Parsova KE, Baştopçu M, Karataş M, Yelgeç NS. Does premature ventricular complex impair left ventricular diastolic functions?. Anatol J Cardiol. 2023;27(4):217-222.
2. Ito S, Tada H, Naito S. Development and validation of an ECG algorithm for identifying the optimal ablation site for idiopathic ventricular outflow tract tachycardia. J Cardiovasc Electrophysiol. 2003;14(12):1280-1286.
3. Betensky BP, Park RE, Marchlinski FE. The V2 transition ratio: a new electrocardiographic criterion for distinguishing left from right ventricular outflow tract tachycardia origin. J Am Coll Cardiol. 2011;57(22):2255-2262.
4. Yoshida N, Inden Y, Uchikawa T. Novel transitional zone index allows more accurate differentiation between idiopathic right ventricular outflow tract and aortic sinus cusp ventricular arrhythmias. Heart Rhythm. 2011;8(3):349-356.
5. Cheng Z, Cheng K, Deng H. The R-wave deflection interval in lead V3 combining with R-wave amplitude index in lead V1: a new surface ECG algorithm for distinguishing left from right ventricular outflow tract tachycardia origin in patients with transitional lead at V3. Int J Cardiol. 2013;168(2):1342-1348.
6. Yoshida N, Yamada T, McElderry HT. A novel electrocardiographic criterion for differentiating a left from right ventricular outflow tract tachycardia origin: the V2S/V3R index. J Cardiovasc Electrophysiol. 2014;25(7):747-753.
7. He Z, Liu M, Yu M. An electrocardiographic diagnostic model for differentiating left from right ventricular outflow tract tachycardia origin. J Cardiovasc Electrophysiol. 2018;29(6):908-915.
8. Kaypakli O, Koca H, Sahin DY, Karataş F, Ozbicer S, Koç M. S-R difference in V1-V2 is a novel criterion for differentiating the left from right ventricular outflow tract arrhythmias. Ann Noninvasive Electrocardiol. 2018;23(3):-.
9. Bozyel S, Şimşek E, Koçyiğit Burunkaya D. Artificial intelligence-based clinical decision support systems in cardiovascular diseases. Anatol J Cardiol. 2024;28(2):74-86.
10. Hancock EW, Deal BJ, Mirvis DM. American College of Cardiology Foundation; Heart Rhythm Society. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram: part V. J Am Coll Cardiol. 2009;53(11):992-1002.
11. Okuyan E. Big data in cardiology. Anatol J Cardiol. 2019;22(2):23-24.
12. Patel S, Kwak L, Agarwal SK. Counterclockwise and clockwise rotation of QRS transitional zone: prospective correlates of change and time-varying associations with cardiovascular outcomes. J Am Heart Assoc. 2017;6(11):-.
13. Kırboğa KK, Küçüksille EU. Identifying cardiovascular disease risk factors in adults with explainable artificial intelligence. Anatol J Cardiol. 2023;27(11):657-663.
14. Moreno-Sánchez PA, García-Isla G, Corino VDA. ECG-based data-driven solutions for diagnosis and prognosis of cardiovascular diseases: a systematic review. Comput Biol Med. 2024;172():-.
15. Zheng J, Fu G, Abudayyeh I. A high-precision machine learning algorithm to classify left and right outflow tract ventricular tachycardia. Front Physiol. 2021;12():-.
16. van de Leur RR, de Brouwer R, Bleijendaal H. ECG-only explainable deep learning algorithm predicts the risk for malignant ventricular arrhythmia in phospholamban cardiomyopathy. Heart Rhythm. 2024;21(7):1102-1112.
17. Bhatia NK, Koscova Z, Rogers AJ. ID: 4120879 Deep learning approaches for prediction of ventricular arrhythmias using upstream electrograms from intracardiac devices. Heart Rhythm. 2024;21(9):S774-.