What question did this study set out to answer?

This research aims to develop an interpretable deep learning framework for accurate multi-label arrhythmia classification from ECG signals.

February 28, 2026Scientific Reports0 citationsOpen Access

A joint CNN-Bi-LSTM-transformer architecture with SHAP explanations for multi-label arrhythmia detection from 12-lead ECGs

Q: What is the clinical evidence from this study?

Study design: Other. Population: Cardiac arrhythmias (n=43101). Intervention: Joint CNN-Bi-LSTM-transformer architecture. Primary outcome: Arrhythmia classification accuracy.

Q: What does this research mean for the field?

A joint CNN-Bi-LSTM-transformer architecture achieves an accuracy of 91.24% and a macro-F1 score of 83.01% for multi-label classification of 27 arrhythmia classes from 12-lead ECGs. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

MAMohammed T. Al-Bairmani MYM. Yazdchi FNFahimeh Nasimi

Key Result

A joint CNN-Bi-LSTM-transformer architecture achieved an accuracy of 91.24% and a macro-F1 score of 83.01% for multi-label classification of 27 arrhythmia classes from 12-lead ECGs.

Key Points

This research aims to develop an interpretable deep learning framework for accurate multi-label arrhythmia classification from ECG signals.
Developed a joint CNN-Bi-LSTM-transformer model to process ECG data.
Trained on the PhysioNet/Computing in Cardiology Challenge 2020 dataset with over 43,000 ECG recordings.
Utilized SHAP for interpretability, attributing predictions to specific ECG waveform segments.
Achieved 91.24% accuracy and 83.01% macro-F1 score in arrhythmia classification.
Exhibited an AUC exceeding 0.98 for life-threatening arrhythmias like VPB and AF.
Demonstrated computational efficiency with <2 MB footprint and <10 ms inference latency.

Structured PICO

Does a joint CNN-Bi-LSTM-transformer architecture with SHAP explanations improve multi-label arrhythmia detection accuracy in 12-lead ECG recordings?

Population

43,101 expert-annotated multi-label 12-lead ECG recordings across 27 arrhythmia classes from the PhysioNet/Computing in Cardiology Challenge 2020 dataset. Mean age approximately 60 years, 53% male, 47% female.

Intervention

Joint CNN-Bi-LSTM-transformer architecture with SHAP (SHAPley Additive exPlanations) for multi-label arrhythmia detection

Outcome

Diagnostic performance measured by accuracy, macro-F1 score, and Area Under the ROC Curve (AUC)surrogate

A novel deep learning architecture combining CNN, Bi-LSTM, and transformer layers with SHAP explanations provides highly accurate, computationally efficient, and clinically interpretable multi-label arrhythmia detection from 12-lead ECGs.

Abstract

Interpretable, automated Artificial Intelligence (AI) solutions are essential for accurate 12-lead electrocardiogram (ECG) arrhythmia classification because they remove the time-consuming and inconsistent aspects of manual interpretation. Current models are limited in complexity, data variety, and validation. This paper proposes a novel Deep Learning (DL) architecture that combines Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (Bi-LSTMs), and transformer layers to jointly extract morphological, temporal, and spatial patterns from ECG signals. The model was trained and evaluated on the PhysioNet/Computing in Cardiology Challenge 2020 dataset, comprising more than 43, 000 multi-label ECG recordings across 27 arrhythmia classes. It achieved an accuracy of \: 91. 24\%, a macro-F1 score of \: 83. 01\%, and an Area Under the ROC Curve (AUC) exceeding \: 0. 98 for life-threatening arrhythmias such as Ventricular Premature Beats (VPB) and Atrial Fibrillation (AF). To ensure clinical transparency, the model integrates SHAP (SHAPley Additive exPlanations), enabling case-by-case interpretability by attributing predictions to physiologically relevant waveform segments and ECG leads. This approach aligns with cardiologists’ diagnostic reasoning and supports real-world decision-making. Additionally, the model is computationally efficient, with a footprint of \: <2\: MB and inference latency of \: <10\: ms, enabling deployment in telemedicine, wearable monitoring systems, and critical care settings. The proposed framework achieves high diagnostic performance, robustness to class imbalance, and human-level interpretability simultaneously, providing a reliable, scalable solution for automated ECG analysis. These findings advance the application of explainable DL algorithms in cardiovascular diagnostics.

Ask AI

Helpful

Bookmark

View Full Paper