[Research on arrhythmia classification algorithm based on adaptive multi-feature fusion network].

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Context:

• Arrhythmia classification from electrocardiogram (ECG) data is clinically significant for early disease screening.
• Effective feature selection under limited abnormal sample supervision remains a challenge.
• Deep learning offers automated analysis and rapid classification of ECG signals.

Purpose:

• To propose an adaptive multi-feature fusion network for enhanced arrhythmia classification.
• To effectively extract and fuse time-domain and frequency-domain ECG features.
• To address the challenge of limited abnormal sample supervision in arrhythmia detection.

Summary:

• The proposed algorithm utilizes a 1D-CNN for time-domain features (RR intervals) and MFCC with a 2D-CNN for frequency-domain features.
• An adaptive weighting strategy fuses these extracted deep features for robust arrhythmia classification.
• The MIT-BIH arrhythmia database was used for evaluation under an inter-patient paradigm.

Impact:

• The algorithm achieved high classification accuracy with an average F1-score of 71.3%.
• Demonstrates potential for algorithmic support in arrhythmia classification for wearable devices.
• Provides a valuable tool for early screening and management of cardiac arrhythmias.