Unsupervised feature learning for electrocardiogram data using the convolutional variational autoencoder.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Traditional electrocardiogram (ECG) feature extraction relies heavily on manual, rule-based methods, which are complex and may miss subtle diagnostic indicators.
• The difficulty in defining all relevant ECG features manually necessitates advanced automated approaches for robust analysis.

Purpose of the Study:

• To propose and validate an unsupervised feature learning method using a convolutional variational autoencoder (CVAE) for extracting ECG features from unlabeled data.
• To demonstrate the effectiveness of CVAE-derived features in characterizing diverse ECG rhythms and improving arrhythmia classification through transfer learning.

Main Methods:

• Trained a CVAE on a large dataset of 596,000 ECG samples from intensive care unit patients.
• Validated CVAE features through clustering, latent space exploration, and anomaly detection on external datasets.
• Applied CVAE features and weights for transfer learning in classifying 12 types of arrhythmias using extreme gradient boosting.

Main Results:

• CVAE features successfully reflected various ECG rhythms without additional training.
• Achieved an f1-score of 0.86 for arrhythmia classification using only CVAE features with extreme gradient boosting.
• Transfer learning with CVAE encoder weight initialization improved model performance by 5% compared to random initialization.

Conclusions:

• Unsupervised feature learning with CVAE offers a powerful alternative to traditional feature extraction for ECG analysis.
• CVAE can effectively extract meaningful ECG characteristics from unlabeled data, facilitating downstream tasks like arrhythmia detection.