Privacy-Preserving Methods for Vertically Partitioned Incomplete Data.

Area of Science:

• Health Informatics
• Data Science
• Privacy-Preserving Technologies

Background:

• Distributed health data networks are increasingly common but face challenges with missing data.
• Current methods often require centralizing data, posing privacy risks and limiting collaboration.
• Vertically partitioned data, where different institutions hold different attributes for the same subjects, is a common scenario.

Purpose of the Study:

• To propose a novel privacy-preserving distributed analysis framework for handling missing data in vertically partitioned health datasets.
• To enable robust data analysis without compromising individual patient privacy.
• To overcome the limitations of traditional data pooling methods in distributed networks.

Main Methods:

• Developed a framework where each institution computes local statistics from private data.
• Intermediate aggregated statistics are shared to construct a global model for missing data imputation.
• The framework is designed for vertically partitioned datasets, ensuring no individual-level data is shared.

Main Results:

• Simulation studies show the proposed privacy-preserving methods perform comparably to pooled data methods.
• The framework significantly outperforms naive methods for handling missing data.
• Analysis of a real-world dataset demonstrates the practical applicability of the proposed methods.

Conclusions:

• The proposed framework offers a privacy-preserving solution for missing data in distributed health networks.
• It facilitates cross-institutional collaboration by eliminating the need for data pooling.
• This approach enhances public trust by ensuring the security of sensitive health information.