DPFact

DPFact performs privacy-preserving distributed tensor factorization on electronic health records to derive clinically meaningful phenotypes for computational phenotyping.

Key Features:

• Privacy-Preserving Mechanisms: Incorporates differential privacy by sharing differentially private intermediary results to prevent exposure of sensitive patient information.
• Collaborative Learning Framework: Implements a distributed approach where hospitals perform local tensor factorization and periodically communicate differentially private intermediary results to a semi-trusted server for aggregation.
• Structured Sparsity Term: Integrates a structured sparsity term to handle heterogeneous patient populations and improve phenotype robustness across diverse datasets.
• Output Perturbation Technique: Applies output perturbation by adding noise to updated intermediary results sent every several iterations after local tensor decomposition to obfuscate sensitive information.
• Performance Evaluation: Demonstrated on real-world and synthetic datasets to outperform state-of-the-art baseline methods in accuracy and communication efficiency under stringent privacy constraints.

Scientific Applications:

• Computational Phenotyping: Transforms large-scale EHR data into concise clinical concepts and phenotypes for research and clinical insight.
• Multi-institutional Collaborative Studies: Enables cross-hospital phenotype discovery and collaborative research without sharing raw EHR data.

Methodology:

Performs distributed tensor factorization with local tensor decomposition at each hospital, periodic transmission of differentially private intermediary results to a semi-trusted server for aggregation, inclusion of a structured sparsity term, and output perturbation via added noise every several iterations.