Improving Automatic Detection of ECG Abnormality with Less Manual Annotations using Siamese Network.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Electrocardiography (ECG) is a crucial non-invasive diagnostic tool.
• Automatic interpretation algorithms are increasingly used for ECG analysis.
• ECG datasets require extensive expert annotations, which are costly and difficult to obtain.

Purpose of the Study:

• To develop a weakly supervised pretraining method for ECG abnormality detection.
• To leverage physician-written diagnostic information to improve feature representation of ECG signals.
• To reduce the reliance on expert annotations in ECG analysis.

Main Methods:

• A Siamese neural network-based weakly supervised pretraining method was proposed.
• The method utilized original diagnostic text from physicians.
• This approach generates useful feature representations from ECG signals.

Main Results:

• The proposed method significantly improved ECG abnormality detection performance.
• Models trained with 1/8 annotated ECG data outperformed fully annotated classical models.
• A substantial reduction in annotation resource requirements was demonstrated.

Conclusions:

• Weakly supervised pretraining using physician notes is effective for ECG abnormality detection.
• This method offers a way to save significant annotation resources.
• The framework is adaptable to other ECG-related tasks with appropriate text similarity metrics.