A Markov Chain Replacement Strategy for Surrogate Identifiers: Minimizing Re-Identification Risk While Preserving...

Area of Science:

• Health Informatics
• Data Privacy
• Machine Learning

Background:

• Personal Health Information (PHI) requires robust protection against re-identification.
• Existing Hiding in Plain Sight (HIPS) strategies use surrogate values to replace PHI.
• Evaluating the effectiveness and utility of different HIPS strategies is crucial for data security.

Purpose of the Study:

• To compare the privacy-preserving benefits and information extraction utility of three HIPS strategies: Consistent, Random, and a novel Markov model.
• To assess these strategies on both simulated and real-world clinical data across various false negative error rates (FNER).

Main Methods:

• Implemented and evaluated three HIPS strategies: Consistent, Random, and Markov model-based substitution.
• Tested strategies on simulated PHI distributions and real clinical corpora from two institutions.
• Quantified PHI leakage and information extraction utility using a range of FNERs (0.1% to 5%).

Main Results:

• The Markov model strategy consistently outperformed Consistent and Random strategies in reducing PHI leakage.
• PHI leakage was reduced from 27.1% to 0.1% at 0.1% FNER and 94.2% to 57.7% at 0.5% FNER using the Markov strategy compared to Consistent substitution.
• Modern deep learning methods showed similar performance across strategies, while older machine learning techniques were impacted by context changes.

Conclusions:

• The Markov surrogate generation strategy offers substantial improvements in reducing inadvertent PHI release.
• This approach balances enhanced privacy with the utility of clinical data for information extraction.
• The Markov strategy represents a significant advancement in secure handling of sensitive health information.