A neural network model based on attention pooling and adaptive multi-level feature fusion for arrhythmia automatic...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Cardiovascular disease incidence is rising, making timely arrhythmia detection critical.
• Electrocardiogram (ECG) is vital for diagnosing and monitoring heart health.
• Current inter-patient automated arrhythmia detection faces challenges with manual features and detecting anomalies.

Purpose of the Study:

• To propose a neural network model enhancing automatic detection of abnormal arrhythmia categories in the inter-patient paradigm.
• To improve performance by focusing on key ECG features and addressing signal scale differences.

Main Methods:

• Developed a neural network model incorporating Attention Pooling (AP) and Adaptive Multilevel Feature Fusion (AMFF).
• AP focuses on crucial channel and spatial features, reducing redundant information.
• AMFF adaptively fuses multilevel ECG features to enhance model expression capability.

Main Results:

• Achieved 99.32% accuracy in the intra-patient paradigm and 93.35% in the inter-patient paradigm (AAMI criteria, MIT-BIH database).
• Demonstrated strong performance in N-category classification and anomaly detection (S, V, F) within the inter-patient paradigm.
• The model exhibits balanced performance across different arrhythmia categories.

Conclusions:

• The proposed AP and AMFF neural network model significantly enhances inter-patient arrhythmia detection performance.
• The model effectively addresses challenges in detecting abnormal categories and ECG signal scale variations.
• This approach offers a more robust and balanced solution for automated arrhythmia diagnosis.