New AI model protects patient privacy in ECG data

A common misperception exists that electrocardiograms (ECGs) simply contain data about heart activity. However, modern ECGs enhanced with artificial intelligence (AI) can contain data about a patient's sex, age, race and even exact identity derived from ECG signals, raising fresh privacy concerns.

To address these worries, researchers from the University of Kansas have developed a privacy-preserving AI model called (PP-VAE) to protect personally sensitive data.

"Modern AI systems may infer sensitive traits from ECG signals, including approximate age ranges and other personal soft-biometric from the signals," said Fairuz Shadmani Shishir, a doctoral student in electrical engineering & computer science at KU who led the study. "Our goal was to develop a method that preserves clinically useful information in ECGs while reducing the exposure of sensitive personal attributes such as age, sex and demographic details."

Shishir and a KU Medical Center research team detail their new method in a recent issue of Scientific Reports.

"We proposed an AI-driven model that analyzes ECG signals to predict clinically important outcomes such as left ventricular ejection fraction (LVEF), which is an indicator of heart abnormalities and early mortality risk," he said. "At the same time, the model is designed to reduce the exposure of sensitive biometric information, including age, sex and demographic characteristics derived from ECG signals."

Shishir said the research was necessary because in the health care sector, companies and medical institutions often share ECGs and other health information across organizations.

"Protecting patient privacy is essential when sharing medical data," he said. "Our goal was to enable secure sharing of clinically useful ECG information without unnecessarily exposing sensitive personal attributes."

Shishir and his collaborators from KU Medical Center reduced identifiability of soft biometrics using independent convolutional neural networks models while retaining clinically useful predictions such as left ventricular hypertrophy and five-year mortality.

"These findings demonstrate the effectiveness of our approach for retaining ECG data while protecting patient privacy," the authors wrote.

Shishir's co-authors included Sumaiya Shomaji, assistant professor in the Department of Electrical Engineering & Computer Science at KU; Amit Noheria, associate professor in the Department of Cardiovascular Medicine at KU Medical Center; and Christopher Harvey and Amulya Gupta from the Department of Cardiovascular Medicine at KU Medical Center.

The researchers argue their method could help hospitals and research institutions share ECG data safely, allowing collaboration and AI development without compromising patient privacy.

"In our experiments, we compared our model's performance to other state-of-the-art models," Shishir said. "We demonstrated that our model has competitive performance compared with other machine-learning approaches. The model performs well in predicting heart disease and early mortality risk while revealing less biometric information from ECG signals."

Further, the KU innovation could help reduce bias in medical care that can lead to underdiagnosis and undertreatment of marginalized groups and women.

"Bias is an important issue to address," Shishir said. "In our models, we aimed to include balanced proportions of male and female patients, as well as balanced representation among white, nonwhite and other racial groups. This was one way we attempted to minimize bias. At the same time, our model was trained using data from KUMC, and although we validated it on public datasets, future work will involve training on datasets from different regions around the world. This will help us better evaluate bias and improve the model's ability to generalize across populations."

Before the technology can be widely adopted, the researchers said building trust and accessibility would be vital.

"We believe there are two key reasons people should use it," Shishir said. "First, the model is designed to generalize across patients in the United States. Second, we plan to make the model publicly available so that anyone can use it. Making the model publicly available follows common practices in the AI field. Institutions will be able to use our model and potentially build their own versions trained on their own datasets. Our goal is to release the model publicly in the future."

The researchers also acknowledged the American Heart Association in supporting this work.