New AI model protects patient privacy in ECG data

There is a common misconception that an electrocardiogram (ECG) only contains data about the heart’s activity. However, modern ECGs enhanced with artificial intelligence (AI) can include data about a patient’s gender, age, race, and even precise identity derived from the ECG signal, raising new privacy concerns.

To address these concerns, researchers at the University of Kansas developed a privacy-preserving AI model called Protect Personally Sensitive Data (PP-VAE).

“Modern AI systems have the potential to infer sensitive characteristics from ECG signals, including approximate age range and other soft biometrics of the individual,” said Fairuz Shadmani Shishir, a KU electrical engineering and computer science doctoral student who led the study. “Our goal was to develop a way to preserve clinically useful information in the ECG while reducing the exposure of sensitive personal attributes such as age, gender, and demographic details.”

Shishir and his research team at KU Medical Center detail their new method in the current issue. Scientific report.

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

Shishir said the study was necessary because in the medical field, companies and medical institutions often share electrocardiograms and other health information between organizations.

“Protecting patient privacy is essential when sharing health data,” he said. “Our goal was to enable the secure sharing of clinically useful ECG information without needlessly exposing sensitive personal attributes.”

Shishir and colleagues at KU Medical Center used an independent convolutional neural network model to reduce the discriminability of soft biometrics while preserving clinically useful predictions such as left ventricular hypertrophy and 5-year mortality.

“These findings demonstrate the effectiveness of our approach for preserving ECG data while protecting patient privacy,” the authors wrote.

Shishir’s co-authors include Sumaiya Shomaji, an assistant professor in KU’s Department of Electrical Engineering and Computer Science. Amit Noheria, associate professor of cardiovascular medicine at KU Medical Center. Christopher Harvey and Amulya Gupta, Department of Cardiovascular Medicine, KU Medical Center;

The researchers claim their method can help hospitals and research institutions securely share EKG data, enabling collaboration and AI development without compromising patient privacy.

“In our experiments, we compared the performance of our model with other state-of-the-art models,” Shishir said. “We demonstrated that our model has competitive performance compared to other machine learning approaches. While the model reveals less biological information from the ECG signal, it performs better in predicting heart disease and early death risk.”

Additionally, KU’s innovations could help reduce bias in healthcare that can lead to underdiagnosis and undertreatment of marginalized groups and women.

“Stigma is an important issue that needs to be addressed,” Shishir said. “In our model, we aimed to include a balanced proportion of male and female patients, and a balanced representation between white, non-white, and other racial groups. This was one way to minimize bias. At the same time, our model Although we trained it using our data and validated it on public datasets, future work will include training on datasets from different regions around the world. This will allow us to better assess bias and improve the model’s ability to generalize across populations.”

The researchers said building trust and accessibility is essential before the technology is widely adopted.

“We think there are two main reasons people should use it,” Shishir says. “First, the model is designed to be generalizable across patients in the United States. Second, we plan to make the model publicly available for anyone to use. Making the model publicly available follows common practice in the AI ​​field. Facilities may use our model to build their own versions trained on their own datasets. Our goal is to make the model publicly available in the future.”

The researchers also acknowledged that the American Heart Association supported the study.