Synthetic electronic health records generated with variational graph autoencoders

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Data Science

Background:

• Patient privacy concerns impede healthcare software development and AI adoption.
• Limited data sharing between healthcare organizations results in unrepresentative patient cohorts and poor statistical models.
• Synthetic data offers a solution to overcome data scarcity in healthcare.

Purpose of the Study:

• To develop a generative neural network model for creating realistic synthetic health records.
• To address the challenge of data privacy in healthcare data sharing.
• To enable the safe and effective use of artificial intelligence in healthcare.

Main Methods:

• Utilized deep neural network architectures, specifically a variational graph autoencoder (VGAE).
• Generated synthetic patient trajectories represented as linear-sequence graphs of clinical events.
• Trained the model on real-world electronic health records to learn statistical properties.

Main Results:

• The model successfully generated synthetic health records with realistic clinical timelines.
• Generated data exhibited the same statistical properties as the original training data.
• The synthetic patient trajectories were realistic and preserved patient privacy.

Conclusions:

• The proposed generative neural network model effectively creates realistic synthetic health records.
• This approach overcomes privacy barriers, facilitating secure data sharing across healthcare organizations.
• The technology supports the advancement and integration of artificial intelligence in healthcare delivery.