Enhancing Electrocardiogram Classification with Multiple Datasets and Distant Transfer Learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Electrocardiogram (ECG) classification is vital for diagnosing cardiovascular diseases and heart conditions.
• Small ECG datasets pose a significant challenge for deep learning models, limiting classification performance.
• Traditional transfer learning is effective but restricted by the need for similar source domains.

Purpose of the Study:

• To enhance ECG classification accuracy using distant transfer learning.
• To propose a novel algorithm, GANAD-DFCNTA, to bridge knowledge transfer from distant domains.
• To overcome the limitations of small ECG datasets in deep learning models.

Main Methods:

• Developed a generative adversarial network-based auxiliary domain with domain-feature-classifier negative-transfer-avoidance (GANAD-DFCNTA) algorithm.
• Utilized eight benchmark datasets: four ECG datasets and four from distant domains (ImageNet, COCO, WordNet, Sentiment140).
• Evaluated the algorithm's performance in transferring knowledge from distant sources to ECG classification tasks.

Main Results:

• The proposed GANAD-DFCNTA algorithm achieved an average accuracy improvement of 3.67% to 4.89% on ECG classification.
• Compared to traditional transfer learning, the algorithm showed an average accuracy improvement of 0.303% to 5.19%.
• Ablation studies validated the effectiveness of individual components within the GANAD-DFCNTA algorithm.

Conclusions:

• Distant transfer learning, facilitated by GANAD-DFCNTA, effectively enhances ECG classification performance.
• The proposed method successfully bridges knowledge transfer from diverse, distant domains to ECG data.
• GANAD-DFCNTA offers a promising solution for improving deep learning model performance on small ECG datasets.