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

Area of Science:

• Cardiovascular Diagnostics
• Artificial Intelligence in Medicine
• Deep Learning for Signal Processing

Background:

• Manual interpretation of 12-lead electrocardiograms (ECG) for arrhythmia classification is time-consuming and inconsistent.
• Existing automated AI models often lack complexity, data variety, and robust validation.
• There is a need for interpretable AI solutions to improve accuracy and clinical trust in ECG analysis.

Purpose of the Study:

• To develop a novel Deep Learning (DL) architecture for accurate and interpretable ECG arrhythmia classification.
• To combine Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (Bi-LSTMs), and transformer layers for comprehensive ECG pattern extraction.
• To ensure clinical transparency and support diagnostic reasoning through explainable AI (XAI) methods.

Main Methods:

• A hybrid DL model integrating CNNs, Bi-LSTMs, and transformer layers was designed.
• The model was trained and validated on the large-scale PhysioNet/Computing in Cardiology Challenge 2020 dataset (over 43,000 ECGs, 27 arrhythmia classes).
• SHAPley Additive exPlanations (SHAP) was incorporated for case-by-case interpretability, linking predictions to ECG features.

Main Results:

• The model achieved high accuracy, a macro-F1 score, and an Area Under the ROC Curve (AUC) exceeding [Formula: see text] for critical arrhythmias like Ventricular Premature Beats (VPB) and Atrial Fibrillation (AF).
• SHAP analysis provided physiologically relevant explanations, aligning predictions with cardiologists' diagnostic reasoning.
• The model demonstrated computational efficiency with a small footprint ([Formula: see text]) and low inference latency ([Formula: see text]).

Conclusions:

• The proposed DL framework offers a robust, scalable, and interpretable solution for automated ECG arrhythmia classification.
• It achieves high diagnostic performance, addresses class imbalance, and provides human-level interpretability.
• This work advances the use of explainable AI in cardiovascular diagnostics, enabling potential deployment in real-time monitoring and telemedicine.