IGAMT: Privacy-Preserving Electronic Health Record Synthesization with Heterogeneity and Irregularity.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Electronic health records (EHR) data are crucial for machine learning (ML) in clinical research.
• Privacy regulations limit access to real EHR data, hindering research.
• Generating synthetic EHR data is a promising solution to privacy concerns.

Purpose of the Study:

• To propose IGAMT, a framework for generating high-quality, privacy-preserved synthetic EHR data.
• To address challenges in EHR data synthesis, including feature heterogeneity, missing values, and irregular temporal measures.
• To achieve differential privacy with an improved privacy-utility trade-off.

Main Methods:

• Developed the IGAMT framework using deep learning for synthetic EHR data generation.
• Incorporated techniques to handle heterogeneous features, missing data, and irregular temporal data.
• Implemented differential privacy mechanisms to ensure data privacy.

Main Results:

• IGAMT demonstrated superior performance compared to baseline and state-of-the-art models.
• Synthetic data generated by IGAMT closely resembles real EHR data.
• Downstream ML applications using IGAMT synthetic data showed improved performance.
• Ablation studies confirmed the effectiveness of IGAMT's components.

Conclusions:

• IGAMT offers a robust solution for privacy-preserving synthetic EHR data generation.
• The framework successfully balances data utility and privacy.
• IGAMT advances the potential of ML-driven clinical research using synthetic EHR data.