Interpretable Feature Generation in ECG Using a Variational Autoencoder

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiogram (ECG) signal analysis is crucial for diagnosing cardiovascular diseases.
• Limited availability of labeled ECG data hinders the training of robust machine learning models.
• Feature extraction from ECG signals is essential for efficient data representation and interpretation.

Purpose of the Study:

• To propose a novel method for generating single cardiac cycle ECG signals using variational autoencoders.
• To develop a concise feature vector for ECG signal encoding.
• To enhance the quality of automatic cardiovascular disease diagnostics and address data limitations for supervised learning.

Main Methods:

• Utilized a variational autoencoder (VAE) for ECG signal generation.
• Encoded original ECG signals into a reduced set of features.
• Extracted a new vector of 25 interpretable features.

Main Results:

• Generated ECG signals exhibit a natural appearance.
• Achieved a low Maximum Mean Discrepancy (MMD) value of 3.83 × 10^-3, indicating high-quality generation.
• Successfully extracted 25 new, often interpretable, features from ECG signals.

Conclusions:

• The proposed VAE-based method effectively generates realistic ECG signals.
• The extracted features can significantly improve the accuracy of automated cardiovascular disease diagnostics.
• Synthetic ECG generation using this method can overcome the scarcity of labeled ECG data for supervised learning applications.