A knowledge embedded multimodal pseudo-siamese model for atrial fibrillation detection.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Atrial fibrillation (AF) diagnosis from ECGs is time-consuming and challenging due to noise and limited context.
• Existing ECG classification models struggle with misclassifying noise and ignoring long-term signal information, impacting AF detection performance.

Purpose of the Study:

• To develop an accurate and efficient AI model for atrial fibrillation detection.
• To address limitations of current models, including noise misclassification and lack of long-term context analysis.

Main Methods:

• Proposed a knowledge-embedded multimodal pseudo-siamese model incorporating a temporal-spatial pseudo-siamese network (TSPS-Net) and a knowledge-embedded noise filter.
• Utilized a parallel siamese network architecture for multimodal representation processing and a spatiotemporal collaborative fusion mechanism (STCFM) for feature fusion.
• Integrated manually designed features, informed by medical knowledge, to differentiate noise and enhance deep ECG features.

Main Results:

• The model achieved an average accuracy of 82.17% and 99.11% on the CinC 2017 and MIT-BIH AF datasets.
• The F1 scores were 0.787 for CinC 2017 and 0.970 for the MIT-BIH AF dataset, demonstrating high performance.

Conclusions:

• The proposed knowledge-embedded multimodal pseudo-siamese model effectively improves AF detection accuracy.
• The integration of medical knowledge and advanced deep learning techniques offers a promising solution for accurate and efficient arrhythmia diagnosis.