Probabilistic record linkage of de-identified research datasets with discrepancies using diagnosis codes

Area of Science:

• Health Informatics
• Biostatistics
• Data Science

Background:

• Linking de-identified patient datasets is crucial for health research.
• Existing methods struggle with datasets lacking personal health identifiers.
• Diagnosis codes are often available but contain discrepancies.

Purpose of the Study:

• To develop and evaluate a novel algorithm for probabilistic linkage of de-identified patient datasets using only diagnosis codes.
• To compare the algorithm's performance against the standard Fellegi-Sunter method.
• To provide an open-source implementation for practical application.

Main Methods:

• Bayesian modeling of binarized diagnosis codes.
• Probabilistic linkage at the patient level, considering all data simultaneously.
• Validation through simulation studies and real-world electronic health record data.

Main Results:

• The algorithm demonstrated good performance in both simulation and real-world use cases.
• It favorably compared to the standard Fellegi-Sunter algorithm.
• Achieved accurate patient-level linkage using only diagnosis codes.

Conclusions:

• Probabilistic linkage of de-identified datasets is feasible using only diagnosis codes.
• The proposed Bayesian approach offers a robust alternative to traditional methods.
• An open-source R package ('ludic') is available for broader adoption.