Developing an Incremental Privacy-Preserving Record Linkage Method to Improve Data Sharing and Data Quality

• Download Summary

What was the project about?

Researchers often have trouble collecting complete information on patient health, as patients may receive care at different places. Linking patient records from different places may help researchers get a more complete picture.

One way to link records is through personal information, such as names and birth dates. But this method increases risks to patient privacy. Another way, known as privacy-preserving record linkage, or PPRL, masks personal information. But current PPRL methods only work when linking entire sets of patient data, including data that have already been shared and linked. Linking entire data sets takes a long time. Also, sharing the same records multiple times increases data privacy risks.

In this study, the research team developed and tested a new PPRL method called incremental PPRL. This method links only new or updated data rather than re-linking entire data sets.

What did the research team do?

The research team developed the new PPRL method based on current methods. They then tested the new method by using it to link test data sets they created. Next, the research team used real patient data to look at how well the new method performed compared to two current record linkage methods that link entire data sets.

The real patient data the research team used included records from 2011 to 2013 from five health systems in the Colorado Congenital Heart Disease registry. The team first linked records for 4,940 patients ages 11–64. They carefully reviewed the linked records to see if they were accurate. Then the team linked the same records using the new method and the two current methods. They compared the linked records from the new and current methods with the data set they had reviewed to see how well each method worked.

Patients, a patient representative, and other researchers gave input on the study.

What were the results?

The new method performed as well as the two current methods in linking patient records. All methods accurately linked records about 97 percent of the time.

The research team made their computer program for the new method available online for free.

What were the limits of the project?

Health systems had a hard time pulling only new or updated records from their data sets to use with the new method. Also, when the team used the new method with large data sets, it was less efficient. The research team only tested the new method with data from one state and one health problem.

Future research could test the new method with patient data for other states and health problems.

How can people use the results?

Researchers can reduce privacy risks by using the new method to link new or updated records with existing data sets.

Background

Researchers often have difficulty accessing patients’ entire health histories because health data may be spread over multiple health systems. To improve data completeness and increase data accuracy, researchers link records from different data sources. One method for linking records is to share personally identifiable information, but this method increases risks to patient privacy. Privacy-preserving record linkage (PPRL) methods mitigate this issue by linking encrypted personal information to create patient identifiers that mask personal information. Existing PPRL methods require linking entire sets of patient records instead of only new or updated records. But re-linking entire data sets is inefficient and increases the risk of exposing sensitive patient data.

In this project, the research team wanted to develop a new PPRL method, called incremental PPRL (iPPRL), which links only new or updated records to an existing patient data set.

Objective

(1) To develop and implement a novel iPPRL method; (2) To compare iPPRL with existing linkage methods and validate its accuracy and effectiveness

Study Design

Methods

The research team extended existing PPRL methods to develop a new iPPRL method. The method successively linked incremental data sets to an initial data set; linkage ended when no new data could be added. The team applied the iPPRL method to a simulated data set containing 115,000 records that mimicked real-world data quality issues.

Then, using real patient data, the research team compared the performance of the iPPRL method with two existing methods which require re-linking whole data sets. The team first linked data from five health systems in the Colorado Congenital Heart Disease registry. They manually reviewed the linked records to create a reference data set containing 4,940 linked records. Next the team linked the same records using the iPPRL method and the two existing methods. They compared the linkage results from the iPPRL and existing methods with the reference data set.

Patients, a patient representative, and researchers provided input throughout the study.

Results

The new iPPRL method performed as well as the existing bulk linkage methods. The methods had similar precision (0.99), recall (0.94), and accuracy (0.97).

The research team tested the iPPRL method with synthetic data and made their code available online for free. They incorporated the iPPRL method into an existing web-based record linkage platform called CU Record Linkage.

Limitations

Health systems had challenges extracting incremental data from their systems when the incremental data included both new and existing patients, which was a barrier to implementing the iPPRL method. Large incremental data sets reduced the efficiency of the iPPRL method, making it computationally expensive. The research team tested the iPPRL method using patient data from one state and one health condition. Results may differ with data from other states and health conditions.

Conclusions and Relevance

The new iPPRL method worked as well as existing methods. Researchers can use this method to reduce logistical challenges and protect patient privacy when linking data.

Future Research Needs

Future research could test the method with data from other locations and populations.