Privacy-Preserving AI Protects Sensitive Information in ECG Data

Artificial intelligence applied to electrocardiography can extract more than cardiac rhythm. Algorithms can infer age, sex, race, and even identity from electrocardiogram (ECG) signals, creating privacy and bias risks when data are shared across institutions. Health systems need tools that preserve prognostic value without exposing sensitive attributes. To help address this challenge, researchers have developed a privacy‑preserving ECG model that keeps risk signals while shielding demographic information.

PP‑VAE is a privacy‑preserving artificial intelligence (AI) model developed at the University of Kansas with clinicians at the University of Kansas Medical Center. The system is described in a study published in Scientific Reports. It is designed to protect sensitive attributes in ECG data while sustaining clinically important predictions needed in cardiovascular care.

The model learns ECG representations using a variational autoencoder trained with an adversarial objective to suppress demographic features. The team also used independent convolutional neural network models to curb identifiability of soft biometrics while monitoring preservation of diagnostic content. The approach maintained downstream predictions that matter clinically, including left ventricular ejection fraction (LVEF), left ventricular hypertrophy, and five‑year mortality risk.

Training relied on data from the University of Kansas Medical Center, with validation performed on public datasets. In head‑to‑head comparisons with other state‑of‑the‑art machine‑learning methods, performance remained competitive while disclosing less biometric information from the ECG signal. To mitigate algorithmic bias, the researchers balanced male and female patients and ensured representation across white, nonwhite, and other racial groups during model development.

The team envisions the method enabling hospitals and research institutions to share ECG data for collaboration and algorithm development without compromising patient privacy. They emphasize the need for trust and accessibility, noting plans to release the model publicly so organizations can adapt it to their own datasets. Future work includes training on datasets from additional regions to strengthen generalizability and assess residual bias across diverse populations.

Related Links
University of Kansas Medical Center