ABMA: An Attention-Based Morphology-Aware Framework for Automated 12-Lead ECG Arrhythmia Classification

Area of Science:

• Cardiology and Medical Informatics
• Artificial Intelligence in Healthcare
• Signal Processing and Machine Learning

Background:

• Cardiovascular diseases (CVDs) are a leading global cause of death, necessitating timely diagnosis through methods like electrocardiograms (ECG).
• Manual ECG analysis is time-consuming and requires expertise due to the complexity and overlap of arrhythmia waveforms.
• Existing deep learning (DL) methods often lack attention to the morphology-sensitive temporal structure crucial for accurate arrhythmia diagnosis.

Purpose of the Study:

• To propose an Attention-Based Morphology-Aware (ABMA) framework for enhanced analysis of multilead ECG signals.
• To leverage a hybrid deep learning architecture combining convolutional neural networks (CNNs) and bidirectional long short-term memory (BiLSTM) networks.
• To develop a morphology scoring network (MSN) within the ABMA module to learn both segment and temporal importance for ECG features.

Main Methods:

• Developed a hybrid deep learning architecture integrating multi-scale CNNs for local morphology feature extraction and BiLSTM for temporal rhythm modeling.
• Designed an ABMA module with a morphology scoring network (MSN) to estimate morphology-aware segment importance and temporal feature importance.
• Conducted an extensive ablation study to evaluate the contribution of each component, including the handcrafted feature branch, ABMA module, MSN, and multi-head attention.

Main Results:

• The ABMA framework achieved a mean accuracy of 95.18% ± 1.18% and a weighted F1-score of 95.19% ± 1.18% in classifying sinus rhythm (SA), atrial fibrillation (AFIB), and ventricular tachycardia (VT).
• The complete ABMA framework significantly outperformed the baseline CNN-BiLSTM architecture and all reduced variants in classification performance (p < 0.001).
• Ablation studies confirmed that each architectural component positively contributed to the overall classification accuracy.

Conclusions:

• The proposed ABMA framework demonstrates superior performance in ECG-based arrhythmia classification compared to baseline models.
• The framework effectively utilizes morphology-aware attention mechanisms to focus on critical ECG signal segments.
• While promising, the ABMA framework requires further validation on independent external and prospective clinical datasets to establish generalizability and clinical utility.