Testing Methods to Link Health Data while Protecting Patient Privacy

• Download Summary

What was the project about?

Patients get care at different places, such as doctors’ offices, hospitals, and pharmacies. Linking patients’ data from these places can help researchers get complete information for each patient for use in studying patient health.

Researchers can use two types of methods to help link patient records based on personal data like names and dates of birth. The first type, known as deterministic, is simpler to use. The second type, known as probabilistic, uses complex computer programs that take longer to work and haven’t been studied as much. But when errors exist in the data, probabilistic methods may work better.

In this project, the research team wanted to learn more about the accuracy of deterministic and probabilistic methods for linking patient records.

What did the research team do?

The research team used four data sets of linked records:

• Newborn screening registry records linked to hospital and clinic health records
• Patient records from different hospitals
• Patient records from public health registries
• Patient death records linked to health records

The research team manually reviewed some records from each data set to check if they were linked correctly. Using only the records that were linked correctly, the team created four reference data sets.

Then, the research team used each linkage method to link individual patient records from the reference data sets one more time. The team compared the linkage results from each method to the reference data sets to see which method was more accurate.

Patients, a patient representative, and other community members helped design the study.

What were the results?

Overall, the probabilistic method was more accurate than the deterministic method. It led to better accuracy in all four data sets.

What were the limits of the project?

The research team didn’t check if one method worked better than the other across different racial and ethnic groups.

Future research could look at how the methods work with records from patients of different racial and ethnic backgrounds.

How can people use the results?

Researchers can use probabilistic linkage methods to link patient records.

Background

Patients receive care at different locations, such as doctors’ offices, hospitals, and pharmacies, each of which create and store patient health data differently. To obtain comprehensive information on patient healthcare experiences, researchers can use different methods to link patient records based on individual demographic data. Key-based deterministic methods require a strict match between demographic data fields across data sources to accurately link patient data. Probabilistic methods are more flexible, allowing linkage even when discrepancies exist in individual patient data between data sets. But probabilistic methods have not been tested extensively with real-world healthcare data.

Objective

To evaluate and compare probabilistic record linkage methods and key-based deterministic methods for linking patient data

Study Design

Methods

The research team first identified four data sets of linked records:

• Newborn screening records linked to the Indiana Network for Patient Care (INPC)
• Linked hospital registries that share overlapping patient populations
• Linked public health registry data from public health services and lab data
• Death records linked to INPC

The research team randomly selected linked records from the four data sets. They manually reviewed the linked records to confirm they were accurate. Using only accurately linked records, the team created four gold standard data sets that represented four use cases for evaluating deterministic and probabilistic linkage methods: identifying unscreened newborns, linking records from different hospital registries, deduplicating records in a public health registry, and determining death status.

Then, using deterministic and probabilistic linkage methods, the research team linked the same sets of patient records in each use case. The deterministic method used combinations of encrypted identifiers to protect patient privacy. The probabilistic method used the Fellegi-Sunter algorithm to select fields for linking de-identified records. In each use case, the team compared patient records linked through both methods with the gold standard data sets. They compared the sensitivity, positive predictive values, and F-scores of the two methods in each of the four use cases.

Patients, a patient representative, and community members helped design the study.

Results

Overall, the probabilistic linkage method was more accurate than the deterministic method. The probabilistic method showed better sensitivity and F-scores in all four use cases, with highest scores in the newborn screening and public health registry use cases. The deterministic linkage method produced higher positive predictive values in the newborn screening and public health registry use cases.

Limitations

The probabilistic linkage method requires more computation time and effort. The research team did not evaluate the methods for algorithmic bias that may affect linkage results for different racial and ethnic groups.

Conclusions and Relevance

Incorporating probabilistic methods can improve record linkage accuracy for de-identified data that preserves patient privacy.

Future Research Needs

Future research could continue to study probabilistic linkage results for records from different racial and ethnic groups.