Abstract
This study addresses the critical challenge of sharing sensitive Electronic Medical Record (EMR) data for genomic research while maintaining patient privacy. By leveraging a "Pragmatic Adversary Model," researchers demonstrate that anonymizing small research cohorts against a massive background population (over one million records) preserves 100% of known genetic associations while maintaining provable privacy guarantees.
Introduction
The "Big Data" revolution in medicine relies on Electronic Medical Records (EMRs) to link genetic variants to clinical diseases—a process called a Phenome-Wide Association Study (PheWAS). However, sharing this data is a privacy minefield. Even if names and social security numbers are removed, a patient’s unique combination of diagnosis codes (ICD-9 codes) can act like a fingerprint, allowing an attacker to re-identify them. Traditional anonymization often suppresses rare codes, which can destroy the scientific "signal" necessary for research. This paper asks: Can we use more realistic, pragmatic assumptions about attackers to share high-quality clinical data without compromising patient anonymity?
Methods
The authors employed a technique called $k$-anonymity, where any individual's record is indistinguishable from at least $k-1$ other records. To achieve this, rare diagnosis codes are replaced with broader categories (code generalization) until the privacy threshold is met.
The core innovation is the Pragmatic Adversary Model. Instead of assuming an attacker knows a patient is in a tiny, specific research cohort, the authors assume the attacker only knows the patient is part of the general hospital population. They tested three strategies for a cohort of ~6,000 patients:
- $DEMO_D$: Anonymizing the 6,000 patients in isolation.
- $DEMO_B$: Anonymizing against a medium-sized background (105k records).
- $DEMO_S$: Anonymizing against the entire hospital population (1.3M records).
Results: The Power of the Crowd
The study found that the larger the population you anonymize against, the less you have to "blur" the data. In the 1.3M-person group ($DEMO_S$), rare diseases find "cover" much more easily than in the 6k-person group.
Anonymizing the cohort in isolation ($DEMO_D$) resulted in significant data loss, leading to 13 missed associations and 9 false associations. In contrast, the $DEMO_S$ strategy successfully discovered all 192 genotype-phenotype associations with the same accuracy as the original, non-anonymized data, while retaining nearly 100% of diagnosis counts.
Discussion and Conclusion
This research shows that privacy and scientific utility are not a zero-sum game. By changing the context of anonymization to include a large background population, researchers can maximize both. However, this approach requires access to massive datasets, which may be a barrier for smaller clinics. Future work could explore "federated" background populations to provide similar protection for smaller research sites. While this protects against re-identification via diagnosis codes, the inherent identifiability of raw DNA sequences remains an ongoing challenge in the field.
Summary updated