Exploring the Complexity of Real-World Health Data Record Linkage-An Exemplary Study Linking Cancer Registry and...

Area of Science:

• Health Informatics
• Data Science
• Epidemiology

Background:

• Record linkage is crucial for integrating health data when unique identifiers are absent.
• Quasi-identifiers (e.g., birth year, sex, location) are often used for linkage but can lead to errors.
• Understanding linkage failures is key to improving data quality and research.

Purpose of the Study:

• To examine reasons for linkage failures using quasi-identifiers.
• To develop and evaluate informed linkage algorithms using gold standard data.
• To assess the achievable linkage quality for German health datasets.

Main Methods:

• Utilized patient data from an antidiabetic cohort (German claims) and colorectal cancer registries.
• Applied and compared various linkage algorithms: deterministic, logistic regression, random forests, gradient boosting, and neural networks.
• Performed descriptive analyses to identify causes of linkage discrepancies.

Main Results:

• Gradient boosting achieved the highest performance: 77% precision, 81% recall, and 64% F*-measure.
• Significant portions of patients lacked unique identification via quasi-identifiers: 8% in GePaRD and 33% in cancer registries.
• Identified discrepancies between data sources as a primary reason for linkage failure.

Conclusions:

• Linking German claims and cancer registry data solely on quasi-identifiers yields insufficient quality.
• Leveraging unique identifiers from a subsample to train algorithms is recommended for the entire dataset.
• Gradient boosting demonstrated superior performance in this record linkage task.