Quantifying and improving rheumatoid arthritis algorithm performance in biobank settings.

Area of Science:

• Rheumatology
• Bioinformatics
• Machine Learning

Background:

• Rheumatoid arthritis (RA) diagnosis relies on complex criteria.
• Standard algorithms for identifying RA cases in biobanks often show suboptimal performance.
• Improving RA case identification is crucial for research and clinical applications.

Purpose of the Study:

• To evaluate and enhance the accuracy of rheumatoid arthritis (RA) detection algorithms within a biobank setting.
• To compare the performance of standard RA algorithms against a novel machine learning approach.
• To optimize the identification of RA patients for research cohorts.

Main Methods:

• Retrospective cohort study in the Mayo Clinic Biobank and Tapestry Study.
• RA cases identified using diagnostic codes, anti-cyclic citrullinated peptide (CCP) antibodies, and disease-modifying anti-rheumatic drug (DMARD) prescriptions.
• Manual verification of RA cases by rheumatology physicians.
• Evaluation of rules-based and eMERGE algorithms using positive predictive value (PPV).
• Development of a novel RA algorithm using LASSO-based machine learning with cross-validation.

Main Results:

• 1,316 confirmed RA cases and 82,123 non-RA controls were identified.
• Rules-based algorithms showed varying PPVs (43%-85%) depending on criteria used.
• The eMERGE algorithm achieved a PPV of 77%.
• A novel machine learning algorithm achieved 90% PPV with improved sensitivity (4-11%) over eMERGE.
• Self-reported RA and family history had minimal to negative impact on algorithm performance.

Conclusions:

• Standard RA algorithms perform less effectively in biobank settings compared to administrative data.
• A novel machine learning algorithm demonstrates superior performance in identifying RA cases within a biobank.
• This enhanced algorithm improves sensitivity and accuracy for RA patient cohort identification.