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

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 [Formula: see text], a macro-F1 score of [Formula: see text], and an Area Under the ROC Curve (AUC) exceeding [Formula: see text] 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 [Formula: see text] and inference latency of [Formula: see text], 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.