Learning Explainable Time-Morphology Patterns for Automatic Arrhythmia Classification from Short Single-Lead ECGs.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Daily cardiac health monitoring relies on accurate detection of abnormal heart rhythms like atrial fibrillation (AF).
• Wearable single-lead electrocardiogram (ECG) devices offer a convenient method for continuous cardiac monitoring.
• Classifying variable-length ECG signals presents a challenge for automated analysis.

Purpose of the Study:

• To develop a novel image-based deep learning framework for classifying single-lead ECG recordings into various arrhythmia types.
• To transform 1D ECG signals into 2D time-morphology representations for improved feature learning.
• To enable interpretable pattern recognition and reduce model parameters for efficient arrhythmia classification.

Main Methods:

• A novel image-based deep learning framework was proposed.
• Variable-length 1D ECG signals were converted into fixed-size 2D time-morphology representations.
• A beat-interval-texture convolutional neural network (BIT-CNN) model was employed for classification.

Main Results:

• The BIT-CNN model achieved an overall F1_NAO of 81.75% and F1_NAOP of 76.87% on the PhysioNet/CinC Challenge 2017 dataset.
• The approach provided interpretable time-morphology patterns, aiding visual interpretation.
• Performance was comparable to state-of-the-art methods for variable-length ECG classification.

Conclusions:

• The proposed image-based deep learning framework offers an effective method for classifying arrhythmias from single-lead ECGs.
• This approach facilitates daily cardiac health monitoring through automated detection of abnormal heart rhythms.
• The method demonstrates potential for reducing model complexity while maintaining high classification accuracy.