Privacy-by-design: Case studies in interactive record linkage using a hybrid human-computer system.

Area of Science:

• Health Informatics
• Data Linkage
• Privacy-Preserving Technologies

Background:

• Accurate patient matching across disparate data sources is challenging without a common identifier.
• Interactive record linkage (RL) using hybrid human-computer systems is essential for high-quality patient matching.
• Minimizing information disclosure during RL is critical for patient privacy.

Purpose of the Study:

• To present and evaluate MiNDFIRL (MInimum Necessary Disclosure For Interactive Record Linkage), a hybrid prototype software system.
• To demonstrate MiNDFIRL's ability to maximize linkage accuracy while minimizing information disclosure.
• To assess the effectiveness of MiNDFIRL in real-world case studies.

Main Methods:

• Conducted two user studies involving 10,000 EHR data pairs and 18,240 patient-generated data unique IDs.
• Utilized automated RL followed by manual review by 12 reviewers using MiNDFIRL for human-judgment resolution.
• Employed consensus for resolving reviewer disagreements and conducted semi-structured interviews for system feedback.

Main Results:

• The Random Forest algorithm identified 388 (EHR) and 539 (patient-generated) matches, with an additional 303 and 187 pairs requiring manual review.
• Manual review confirmed 232 (EHR) and 84 (patient-generated) additional true linkages from uncertain pairs.
• MiNDFIRL used only 30% of available identifying information to correctly classify 77% (EHR) and 45% (patient-generated) of uncertain linkages, with names and emails being most frequent.

Conclusions:

• Hybrid human-computer systems with on-demand access and masking reduce disclosure risks in RL.
• MiNDFIRL effectively minimizes false positives and false negatives in real-world patient data linkage.
• Risk quantification and interactive review are key to balancing accuracy and privacy in record linkage.