Effectiveness of Transfer Learning for Deep Learning-Based Electrocardiogram Analysis.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Deep learning models for electrocardiogram (ECG) analysis require large, labeled datasets.
• Data acquisition for biosignal research is challenging for individual researchers.
• Transfer learning offers a potential solution to data limitations in ECG analysis.

Purpose of the Study:

• To investigate the effectiveness of transfer learning for ECG rhythm classification.
• To address the challenge of insufficient training data in deep learning for ECG analysis.
• To enhance the performance of ECG predictive models using transfer learning.

Main Methods:

• Pretrained a convolutional autoencoder (CAE) on 2,648,100 unlabeled ECG samples.
• Applied transfer learning by using pretrained CAE weights to classify 11 ECG rhythms in a dataset of 10,646 ECGs.
• Evaluated model performance using F1-scores and mean squared errors (MSEs) under varying training data sizes and compared with random initialization.

Main Results:

• The convolutional autoencoder (CAE) achieved a mean squared error (MSE) of 626.583.
• Transfer learning resulted in mean F1-scores of 0.857, 0.843, and 0.835 for 100%, 50%, and 25% of the training data, respectively.
• Random initialization yielded significantly lower F1-scores (0.543) at 25% training data compared to transfer learning.

Conclusions:

• Transfer learning effectively mitigates data shortages in deep learning-based ECG analysis.
• The proposed transfer learning approach enhances the accuracy of ECG rhythm classification.
• This method improves the utility of deep learning models in the ECG domain despite limited labeled data.