Generating sequential electronic health records using dual adversarial autoencoder.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Machine Learning

Background:

• Electronic Health Records (EHRs) are crucial for healthcare but raise privacy concerns.
• Existing deep generative models often focus on independent patient visits, not chronological data.
• Synthesizing realistic, sequential EHR data is essential for privacy preservation and downstream applications.

Purpose of the Study:

• To develop a novel deep generative model for synthesizing realistic sequential EHR data.
• To address limitations of existing models in handling chronological clinical information.
• To improve privacy preservation in EHR data synthesis.

Main Methods:

• Proposed a Dual Adversarial Autoencoder (DAAE) model.
• Combined a recurrent autoencoder with two Generative Adversarial Networks (GANs).
• Evaluated DAAE on MIMIC-III and UT Physicians databases for predictive modeling, plausibility, and privacy.

Main Results:

• Generated EHR sequences demonstrated comparable performance to real data in predictive tasks.
• DAAE achieved the highest plausibility scores evaluated by medical experts.
• Differentially private optimization ensured synthetic data generation without increased privacy leakage.

Conclusions:

• DAAE effectively synthesizes sequential EHRs, meeting realism and downstream task reproduction requirements.
• The model addresses key challenges in generating high-quality, privacy-preserving synthetic EHR data.
• DAAE offers a promising approach for leveraging EHR data while mitigating privacy risks.