Cardioattentionnet: advancing ECG beat characterization with a high-accuracy and portable deep learning model

Area of Science:

• Cardiovascular disease research
• Artificial intelligence in medicine
• Signal processing for healthcare

Background:

• Cardiac arrhythmias pose a significant mortality risk and are challenging to diagnose accurately.
• Current diagnostic methods for arrhythmias can lead to misdiagnosis, necessitating improved automated solutions.

Purpose of the Study:

• To develop a portable deep learning model for real-time arrhythmia detection using electrocardiogram (ECG) signals.
• To enhance diagnostic accuracy and patient outcomes through an efficient, automated detection system.

Main Methods:

• Development of CardioAttentionNet (CANet), a deep learning model integrating Bi-directional Long Short-Term Memory (BiLSTM) networks, Multi-head Attention, and Depthwise Separable Convolution.
• Designed for efficient processing of extended ECG patterns and detailed feature extraction on portable devices.

Main Results:

• CANet demonstrated superior predictive performance and stability compared to traditional models in cross-validation.
• Achieved high accuracy in classifying various arrhythmias: Normal (97.37%), Supraventricular (98.09%), Ventricular (92.92%), Atrial Fibrillation (99.07%), and Unclassified (99.68%).
• External evaluations showed an average accuracy of 94.41% with Area Under the Curve (AUC) exceeding 99% for all categories.

Conclusions:

• The proposed deep learning model significantly enhances the accuracy of early arrhythmia diagnosis.
• CANet technology holds potential for continuous, non-invasive monitoring and timely intervention in heart disease patients.