Linking Electronic Health Records for Multiple Sclerosis Research: Comparative Study of Deterministic, Probabilistic,...

Area of Science:

• Health Informatics
• Pharmacoepidemiology
• Data Science

Background:

• Electronic health records (EHRs) are crucial for pharmacoepidemiological research.
• Linking deidentified EHR data across institutions is challenging due to the absence of unique patient identifiers.
• The Saudi Real-World Evidence Network (SRWEN) requires robust data linkage techniques.

Purpose of the Study:

• To evaluate and compare deterministic, probabilistic, and machine learning (ML) data linkage methods.
• To assess linkage performance for deidentified multiple sclerosis (MS) patient data from SRWEN and Ministry of National Guard Health Affairs EHR systems.
• To determine the most effective linkage strategy for real-world evidence generation in MS research.

Main Methods:

• A simulation-based validation framework was employed prior to real-world data linkage.
• Deterministic linkage utilized predefined rules.
• Probabilistic linkage used similarity score-based matching.
• Machine learning approaches included neural networks, logistic regression, and random forest models, employing both similarity score-based and classification methods.
• Performance was evaluated using confusion matrices, focusing on sensitivity, positive predictive value, F1 score, and computational efficiency.

Main Results:

• Deterministic linkage achieved a high simulation F1 score (97.2%) but showed variable real-world match rates (46.6%-86.6%) with high computational efficiency (<1 second/rule).
• Probabilistic linkage demonstrated strong simulation performance (F1 score 93.9%) with higher real-world match rates (65.5%-95.4%) and moderate processing times (0.1-5 seconds/rule).
• Machine learning models achieved the highest F1 scores (up to 99.8%) but were computationally intensive, with highly variable real-world match rates depending on the specific ML approach.

Conclusions:

• Probabilistic linkage offers superior capacity for linking deidentified data, recovering matches missed by deterministic methods.
• This approach is flexible and efficient for real-world scenarios lacking unique identifiers, balancing recall and precision.
• Probabilistic linkage facilitates better integration of diverse data sources, significantly benefiting multiple sclerosis (MS) research.