Privacy-Preserving Tensor Factorization for Collaborative Health Data Analysis.

Area of Science:

• Computational Health Informatics
• Data Science
• Privacy-Preserving Machine Learning

Background:

• Electronic health records (EHRs) contain vast patient data, necessitating efficient computational phenotyping.
• Collaborative analysis of EHRs offers insights but poses privacy risks due to data sharing.
• Existing tensor factorization methods require sharing intermediate results, potentially exposing sensitive patient information.

Purpose of the Study:

• To develop a privacy-preserving collaborative tensor factorization method for computational phenotyping using EHR data.
• To enable hospitals to jointly learn clinical concepts from their private EHR databases without direct data sharing.
• To ensure model interpretability while adhering to rigorous privacy constraints.

Main Methods:

• Proposed DPFact, a method combining differential privacy with collaborative learning for tensor factorization.
• Hospitals locally decompose tensors and share differentially private intermediate results with a semi-trusted server.
• Incorporated a structured sparsity term to address heterogeneous patient populations.

Main Results:

• DPFact demonstrated superior accuracy and communication efficiency compared to state-of-the-art methods on real-world and synthetic datasets.
• The method effectively preserves patient privacy through differential privacy mechanisms.
• Successful collaborative learning of meaningful clinical concepts was achieved.

Conclusions:

• DPFact offers a robust solution for privacy-preserving computational phenotyping using EHR data.
• The method facilitates collaborative analysis across institutions while maintaining stringent data privacy.
• DPFact enhances the accuracy and efficiency of phenotype discovery from distributed EHR datasets.