TL;DR
This paper details the implementation of a privacy-preserving record linkage (PPRL) system across PCORnet, a large network of healthcare data sources. The study quantifies patient overlap between these sources and demonstrates how linking electronic health records (EHRs) with insurance claims significantly enriches patient data, revealing a more complete picture of disease prevalence.
What Problem Does This Paper Address?
Healthcare data is often fragmented across different systems, such as electronic health records (EHRs) within hospitals and administrative claims from insurance providers. While EHRs offer detailed clinical information, they may miss care received outside a specific health system. Conversely, claims data capture a broader range of medically attended events but lack the granular detail of EHRs. This fragmentation creates gaps in understanding a patient's complete health journey. To address this, record linkage is crucial, but sharing direct patient identifiers raises significant privacy concerns. The paper tackles the challenge of implementing and evaluating a privacy-preserving record linkage (PPRL) solution to bridge these data gaps without compromising patient privacy, specifically within the PCORnet network.
Methods
The study employed Datavant's privacy-preserving record linkage (PPRL) solution. This involved transforming personally identifiable information (PII) from participating PCORnet Network Partners into de-identified, encrypted hash tokens. These tokens were designed to be non-reversible and met HIPAA's Privacy Rule standards.
Two main queries were executed:
- Full Population Cohort: This query extracted hash tokens for all patients within each DataMart (a harmonized data repository).
- 2018-2019 Cohort: This query extracted demographic and clinical characteristics for patients with a clinical encounter between January 1, 2018, and December 31, 2019.
The hash tokens were sent to the Duke Clinical Research Institute (DCRI) for matching using Datavant's software, generating a de-identified master patient index. The demographic and clinical characteristics were sent to the Harvard Pilgrim Healthcare Institute (HPHCI). The DCRI then shared the match index (without tokens) with HPHCI to link the datasets and perform analyses.
A key aspect of the methodology was the separation of roles between DCRI and HPHCI to prevent any single entity from having access to both the de-identified tokens and the clinical data, further enhancing privacy. The study also included a token sensitivity analysis to assess the impact of missing Social Security Number (SSN) tokens on matching performance.
Datasets
The study utilized data from 61 PCORnet Network Partners, comprising 59 Clinical Research Networks (CRNs) and 2 Health Plan Research Networks (HPRNs). The data included:
- EHR Data: Harmonized EHR data from CRNs, representing patient encounters within health systems.
- Claims Data: Administrative claims data from HPRNs and some CRNs, representing billed medical events.
The analysis focused on two cohorts:
- Full CDM Cohort: Approximately 170 million patient records.
- 2018-2019 Cohort: Approximately 72.2 million patient records, used for demographic and clinical characteristic analysis.
Key Insights and Ideas
- PPRL is Feasible at Scale: The project successfully implemented a PPRL solution across a large, distributed network of healthcare data, demonstrating its scalability and practicality.
- Overlap is Variable but Significant: Patient overlap between different network partners was not uniform. It was influenced by factors like geographic proximity and organizational relationships, but even low percentages of overlap across large datasets can represent a substantial number of unique individuals.
- Claims Data Dramatically Enhances Data: Linking EHR data with claims data significantly increased the observed prevalence of various diseases and clinical characteristics, highlighting the substantial "information gain" and the limitations of relying on a single data source.
- Token Choice Matters, but SSN Isn't Always King: While SSN-based tokens can be precise, the study found that a combination of other PII (name, gender, DOB) was effective for linkage, and missing SSN data did not significantly hinder matching performance in this context. This is important as SSN collection is declining.
- Timeframe Impacts Overlap: Overlap between data sources is higher when considering longer timeframes, which is a critical consideration for study design.
Key Results
- Overall Patient Records: Approximately 170 million patient records were processed across 61 Network Partners.
- Unique Patients: Approximately 138 million (81%) of these records corresponded to unique patients in the full cohort. In the 2018-2019 cohort, 65.1 million (90.2%) were unique.
- Patient Overlap:
- Full CDM Cohort: 82.1% of unique patients were found in a single partner, and 14.7% were in two partners.
- 2018-2019 Cohort: 90.5% of unique patients were found in a single partner, and 8.5% were in two partners.
- Data Enrichment: Linking EHR with claims data increased disease prevalence by 63% to 173% (relative difference) and 1.9% to 11.7% (absolute difference) for various clinical characteristics.
- Token Sensitivity: The study found that using all available tokens or just a single token (based on name, gender, DOB) resulted in almost identical matching performance, suggesting robustness to missing SSN data.
Limitations and Future Work
- Linkage Validation: The study did not include a direct validation of the linkage algorithm itself by comparing matched charts to confirm identity. While external studies exist, this was not part of the current work.
- Scope of Data Enrichment: Governance limitations restricted the number of variables and the timeframe for analyzing the information gain from claims data. A more comprehensive analysis could reveal even greater insights.
- Future Directions: The authors suggest exploring additional tokens beyond traditional PII, such as cell phone numbers or email addresses, and incorporating historical demographic information to improve linkage over longer periods. They also highlight the need for careful consideration of study design (population, timeframe, partner characteristics) when deciding whether to use record linkage. The findings also suggest that linkage between geographically proximate partners or those with organizational ties is more fruitful.
Summary updated