Explainable Parallel CNN-LSTM Model for Differentiating Ventricular Tachycardia from Supraventricular Tachycardia with Aberrancy in 12-Lead ECGs
Zahra Teimouri-Jervekani, Fahimeh Nasimi, Mohammadreza Yazdchi, Ghazal MogharehZadeh, Javad Tezerji, Farzan Niknejad Mazandarani, Maryam Mohebbi · 2025
Read Explainable Parallel CNN-LSTM Model for Differentiating Ventricular Tachycardia from Supraventricular Tachycardia with Aberrancy in 12-Lead ECGs by Zahra Teimouri-Jervekani, Fahimeh Nasimi, Mohammadreza Yazdchi, Ghazal MogharehZadeh, Javad Tezerji, Farzan Niknejad Mazandarani, Maryam Mohebbi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-14196.txt.
Introduction
Zahra Teimouri-Jervekani and six co-authors build in 2025 a deep learning system to tell apart two racing-heart rhythms that look deceptively alike on an ECG. Ventricular tachycardia can kill, while its mimic, a fast rhythm from above the ventricles gone wide, is treated very differently, so a wrong call carries real danger.
Their model runs a convolutional network and a memory network side by side over twelve-lead tracings, then explains its verdicts so doctors can judge them. The design stresses both accuracy and thrift, aiming at hospital use. The work targets one of emergency cardiology's classic judgment calls.
Catalog record
| CREATOR | Zahra Teimouri-Jervekani, Fahimeh Nasimi, Mohammadreza Yazdchi, Ghazal MogharehZadeh, Javad Tezerji, Farzan Niknejad Mazandarani, Maryam Mohebbi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 39 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-14196
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Explainable Parallel CNN-LSTM Model for Differentiating Ventricular Tachycardia from Supraventricular Tachycardia with Aberrancy in 12-Lead ECGs Zahra Teimouri-Jervekani, Fahimeh Nasimi, Mohammadreza Yazdchi, Ghazal MogharehZadeh, Javad Tezerji, Farzan Niknejad Mazandarani, Maryam Mohebbi Abstract Background and Objective: Differentiating wide complex tachycardia (WCT) is clinically critical yet challenging due to morphological similarities in electrocardiogram (ECG) signals between life-threatening ventricular tachycardia (VT) and supraventricular tachycardia with aberrancy (SVT-A). Misdiagnosis carries fatal risks. We propose a computationally efficient deep learning solution to improve diagnostic accuracy and provide model interpretability for clinical deployment. Methods: A novel lightweight parallel deep architecture is introduced. Each pipeline processes individual ECG leads using two 1D-CNN blocks to extract local features. Feature maps are concatenated across leads, followed by LSTM layers to capture temporal dependencies. Final classification employs fully connected layers. Explainability is achieved via Shapley Additive Explanations (SHAP) for local/global interpretation. The model was evaluated on a 35-subject ECG database using standard performance metrics. Results: The model achieved 95.63 % 95.63\% accuracy ( 95 % 95\% CI: 93.07 – 98.19 % 93.07–98.19\% ), with sensitivity= 95.10 % 95.10\% , specificity= 96.06 % 96.06\% , and F1-score= 95.12 % 95.12\% . It outperformed state-of-the-art methods in both accuracy and computational efficiency, requiring minimal CNN blocks per pipeline. SHAP analysis demonstrated clinically interpretable feature contributions. Conclusions: Our end-to-end framework delivers high-precision WCT classification with minimal computational overhead. The integration of SHAP enhances clinical trust by elucidating decision logic, supporting rapid, informed diagnosis. This approach shows significant promise for real-world ECG analysis tools. Keywords: wide complex tachycardia, ventricular tachycardia, supraventricular tachycardia with aberrancy, explainable artificial intelligence. 1 Introduction The electrocardiogram (ECG) is a crucial biosignal for cardiologists, providing essential information about the heart’s electrical activity. Early detection and discrimination of cardiac arrhythmias can significantly improve patient outcomes and quality of life. One of the main uses of ECG is the differentiation of wide complex tachycardia(WCT). Three primary mechanisms that can lead to WCT are supraventricular tachycardia with intraventricular aberrant conduction, supraventricular tachycardia that conducts to the ventricles via an accessory pathway, and ventricular tachycardia. The primary clinical challenge lies in distinguishing between ventricular tachycardia (VT) and supraventricular tachycardia with aberrancy(SVT-A) [ 1 ] . Figure 1 demonstrates the electrocardiographic overlap between VT and SVT-A, including shared features such as rapid heart rate and widened QRS complexes. This visual comparison reinforces the importance of applying validated algorithms to avoid misdiagnosis in acute settings. WCT should routinely be treated as VT(VT is significantly more common than SVT-A conduction, accounting for over 80\% of cases) unless the diagnosis of SVT-A or SVT with preexcitation is certain [ 2 ] . Adenosine and other atrioventricular(AV) nodal–blocking agents are ineffective and potentially deleterious in patients with VT. Numerous ECG criteria and interpretation algorithms have been developed to help clinicians distinguish VT from SVT-A [ 3 ] , [ 4 ] , [ 5 ] , [ 6 ] , [ 7 ] . Till now the Brugada criteria remain the gold standard due to high sensitivity and guideline endorsement. Brugada et al. introduced a stepwise decision tree algorithm to differentiate between VT and SVT-A. The initial step evaluates all precordial leads (V1–V6) for the presence of RS complexes. VT is diagnosed, if no RS complexes are observed in any lead. If RS complexes are present, the longest RS interval in any precordial lead exceeding 100ms is diagnostic of VT. After these steps, the presence of independent P waves, capture beats, or fusion beats also confirms VT. If the preceding steps are inconclusive, the diagnosis of VT is supported by the presence of a monophasic R, QR, or RS pattern with an R-wave duration greater than 40ms , or a notched S wave in lead V1. Additional supporting criteria include an R/S ratio less than 1 , QS or QR pattern in lead V6, an R-wave onset to S-wave nadir interval greater than 60ms in leads V1 or V2, a notched S wave, or a delayed S-wave nadir (greater than 60ms ), along with a QS or QR pattern in V6 [ 8 ] . This algorithm reportedly achieved a sensitivity of 98.7\% and a specificity of 96.5\% , outperforming traditional criteria. However, because of complex morphological criteria in its final steps, challenges in detecting AV dissociation at high heart rates and interpreter experience further affects accuracy. As a result, there is increasing interest in developing automated, computer-aided techniques for WCT differentiation to support clinicians in patient diagnosis and monitoring [ 9 ] . In recent years, deep learning (DL) algorithms—particularly convolutional neural networks (CNN)—have been widely adopted as automated, computer-aided tools for various biomedical applications, demonstrating promising performance [ 10 ] . DL networks are well-suited for ECG signal classification due to their capability to identify patterns and automatically learn relevant features directly from raw data [ 11 ] . Although several approaches have been introduced for WCT differentiation using deep learning models, many rely on complex, deep architectures that may impede real-time processing. This highlights the need for lightweight DL models capable of delivering accurate WCT classification with minimal computational delay. To counter the practical diagnostic limitations of manual and automated algorithms, as well as supplement their known strengths, we developed and validated a precise, lightweight automated diagnostic system designed to assist cardiologists by offering an interpretable deep learning approach that is time-efficient, cost-effective, reliable, and reduces the incidence of misdiagnosed VTs. The proposed model integrates CNN and recurrent neural networks (RNN), specifically long short-term memory (LSTM) units, to distinguish VT from SVT-A using 12 -lead ECG signals. In summary, the key contributions and innovations of this study are outlined as follows: 1. Localized morphological features were extracted from individual ECG leads using dedicated CNN blocks. These modules employed convolutional layers to capture lead-specific patterns. 2. The derived feature vectors from each lead were concatenated into a unified representation. Subsequently, LSTM networks were applied to model spatial-temporal dependencies across leads, enabling the identification of inter-lead correlations critical for rhythm discrimination. 3. Leave-one-out cross-validation (LOOCV) was rigorously applied to partition training and testing datasets. This approach ensured strict separation of samples across folds, minimizing overfitting risks and preserving the integrity of model performance evaluation. 4. Shapley values are used to show the contribution or the importance of each ECG sample and lead to the prediction of the model. Section 2 first describes the morphology of WCT. Section 3 reviews related work. The proposed system architecture is detailed in Section 4 . Section 5 presents experimental results and compares them with state-of-the-art methods. Finally, Section 6 discusses the current status of the proposed technique, along with its advantages and limitations. Figure 1 : Top) 12-Lead ECG signal with VT. Bottom) 12-Lead ECG signal with SVT-A 2 Morphology of Wide Complex Tachycardia In a healthy heart, electrical impulses typically originate from the sinoatrial (SA) node—commonly referred to as the heart’s natural pacemaker—and subsequently travel to the atrioventricular (AV) node. This electrical conduction triggers ventricular contraction, facilitating the forward movement of blood [ 12 ] . SVT and VT arise from aberrant cardiac signaling. In SVT, altered signaling triggers premature initiation of the heartbeat in the atria, whereas in VT, the premature onset occurs in the ventricles. These disruptions