Estimating the re-identification risk of clinical data sets

Area of Science:

• Health Informatics
• Data Privacy
• Statistical Modeling

Background:

• De-identification of clinical data is essential for protecting patient privacy in research.
• Uniqueness is a key metric for assessing re-identification risk from data linkage attacks.
• Direct measurement of uniqueness is often infeasible, necessitating accurate estimation methods.

Purpose of the Study:

• To evaluate the accuracy of various statistical estimators for measuring uniqueness in clinical datasets.
• To identify the most reliable method for estimating re-identification risk in de-identified health information.

Main Methods:

• Compared four uniqueness estimators (Zayatz, slide negative binomial, Pitman, mu-argus) using Monte Carlo simulations.
• Assessed estimator accuracy on six diverse clinical datasets, varying sampling fractions and population uniqueness.
• Measured median relative error and inter-quartile range across 1000 simulation runs.

Main Results:

• No single estimator demonstrated superior performance across all tested conditions.
• A novel decision rule was developed, selecting between Pitman, slide negative binomial, and Zayatz estimators.
• The developed decision rule exhibited the most consistent median relative error across various datasets and scenarios.

Conclusions:

• An accurate decision rule for estimating uniqueness in clinical data has been identified.
• This rule offers a reliable approach for health privacy researchers and disclosure control professionals.
• Improved estimation of uniqueness enhances the management of re-identification risks in sensitive health data.