Foundation Model-Guided Synthetic EHR Release: Performance Enhancement with Privacy Preservation.

Area of Science:

• Health Informatics
• Machine Learning
• Data Privacy

Background:

• Machine learning on electronic health records (EHR) faces challenges like limited data, privacy concerns, and data distribution shifts.
• Existing synthetic data generators struggle with the small sample sizes common in EHR datasets.

Purpose of the Study:

• To develop a novel synthetic EHR generator, MEDPFN, that addresses limitations of current methods.
• To enable high-utility and privacy-friendly release of EHR data for research and development.

Main Methods:

• MEDPFN is built upon the TabPFN foundation model, incorporating a distribution-aware approach.
• It utilizes TabPFN as a learned compatibility score and adds a Mahalanobis regularizer to stabilize generation from small cohorts.
• The model was trained and evaluated on six public tabular EHR datasets.

Main Results:

• Classifiers trained solely on MEDPFN-generated data achieved performance comparable to or exceeding those trained on real data.
• MEDPFN outperformed several state-of-the-art synthetic data generators across different datasets and models.
• The method successfully generated high-utility synthetic EHR data while preserving privacy.

Conclusions:

• MEDPFN offers a robust solution for generating synthetic EHR data, overcoming limitations of small sample sizes and privacy regulations.
• This approach facilitates the broader use of EHR data for machine learning by enabling secure and effective data sharing.