NoisyFlow: differentially private optimal transport using neural networks for secure biomedical data sharing across...

Area of Science:

• Biomedical data science
• Computational privacy
• Machine learning for healthcare

Background:

• Sharing sensitive patient data across institutions is crucial for improving biomedical models but faces privacy constraints.
• Data distribution shifts (e.g., batch effects) between institutions hinder multi-institutional analyses.
• Existing privacy-preserving methods do not adequately address cross-site distribution mismatch.

Purpose of the Study:

• To develop a privacy-preserving framework for harmonizing multi-institutional biomedical data under distribution shift.
• To enable reliable cross-institutional analyses without compromising patient privacy.
• To address the challenge of integrating data from diverse sources while maintaining data utility.

Main Methods:

• NoisyFlow is a three-stage differentially private framework.
• Stage I: Each site learns a differentially private flow-based generator of its local distribution.
• Stage II: A neural optimal transport map aligns local distributions to a shared reference.
• Stage III: A central server composes models to generate reference-aligned pseudo-data.

Main Results:

• NoisyFlow effectively reduces distribution shift across different biomedical settings.
• The framework preserves downstream analytical utility under formal differential privacy guarantees.
• Demonstrated success in single-cell genomics, histopathology, neurogenomics, and wearable sensing data.

Conclusions:

• NoisyFlow provides a robust solution for privacy-preserving cross-institutional data harmonization.
• The method facilitates reliable multi-institutional analyses by addressing both privacy and distribution shift.
• Enables the development of more generalizable and accurate biomedical models through pooled, harmonized data.