AI-enabled privacy-preserving cardiac diagnostics via electrocardiograms.

Area of Science:

• Cardiovascular Medicine
• Machine Learning
• Biomedical Informatics

Background:

• Electrocardiograms (ECGs) are crucial for cardiovascular health assessment but contain sensitive demographic data.
• Demographic information in ECGs can lead to bias and privacy issues in machine learning models.
• High dimensionality of ECG data complicates the development of fair and private AI.

Purpose of the Study:

• To develop a deep learning framework for learning clinically relevant ECG representations.
• To suppress sensitive demographic information (sex, age, race) within ECG signals.
• To maintain diagnostic accuracy for cardiovascular conditions and mortality prediction.

Main Methods:

• Utilized a variational autoencoder (VAE) with a dual-discriminator architecture.
• Employed adversarial learning to reduce soft biometric encoding (demographics).
• Preserved discrimination of clinically significant features like reduced left ventricular ejection fraction (LVEF).

Main Results:

• Reduced demographic identifiability: AUROC for sex (0.59), age (0.63), race (0.57).
• Maintained clinical prediction accuracy: reduced LVEF (0.82), left ventricular hypertrophy (LVH) (0.72), 5-year mortality (0.67).
• Demonstrated effective privacy preservation while retaining diagnostic utility.

Conclusions:

• The proposed deep learning framework successfully suppresses demographic information in ECGs.
• The method balances patient privacy with the retention of critical clinical diagnostic information.
• This approach offers a promising solution for privacy-preserving analysis of ECG data.