Background/Purpose: To develop and validate a search algorithm with a high specificity and sensitivity to identify rheumatoid arthritis (RA) patients in a large health care system linked by an electronic health record (EHR) system.

Methods: Records from the Medical Archival Retrieval System (MARS) at the University of Pittsburgh Medical Center (UPMC) were used to identify potential RA patients.  We searched the UPMC MARS system for subjects with a 714.0 International Classification of Diseases, 9th revision (ICD-9) code and used a recursive partitioning method to develop a search algorithm for the identification of RA patients; the recursive partitioning method used the 714.0 ICD-9 code and tested 35 additional variables (serology, inflammatory markers, medications and specific words in physician notes).  At each step during the development and validation of the algorithm, patients were classified by the algorithm into those likely or unlikely to have RA and representative sets of these patient records were reviewed to determine whether subjects met ACR/EULAR RA classification criteria.

Results: We initially analyzed the effect of the clinical setting (inpatient vs. outpatient rheumatology clinic) on the identification of RA patients.  For inpatient subjects, there was a low PPV of a 714.0 ICD-9 code for the identification of RA patients (39.0%) whereas for outpatient-rheumatology subjects there was a high PPV of a 714.0 ICD-9 code for the identification of RA patients (87.3%), (n=95, p<0.0001; Fisher’s exact test).  When the records of outpatient-rheumatology patients with and without a 714.0 ICD-9 code were analyzed (N=400), the sensitivity, specificity, PPV and NPV of a 714.0 ICD-9 code was 98%, 88%, 87% and 98%, respectively.  Using recursive partitioning a 3 variable algorithm was identified 1.) 714.0 ICD-9 code, 2.) ratio of the words “rheumatoid arthritis” per rheumatology visit and 3.) ratio of “RA” per rheumatology visit improved the specificity for identifying RA patients.  The sensitivity, specificity, PPV and NPV of the algorithm was 93%, 95%, 94% and 95%, respectively (n=400).  Validation of this algorithm with analysis of an additional 400 subjects produced similar results (95%, 96%, 96%, 95%, respectively).   Using this algorithm to analyze all unique outpatient-rheumatology patients in the 2009 calendar year (n= 5,859) resulted in the identification of 1,495 RA patients.  The final validation step with a sample of these subjects (n=400) resulted in a sensitivity, specificity, PPV and NPV of 93%, 96%, 93% and 96%, respectively. 

Conclusion: The ICD-9 code for RA (714.0) alone was not reliable for identifying RA patients in the inpatient setting and had suboptimal specificity in the outpatient rheumatology setting.  We developed and validated a simple algorithm using recursive partitioning that used 3 variables to identify RA patients.  This simple algorithm, which is now validated for an entire calendar year, represents a substantial improvement in terms of sensitivity and specificity over existing published algorithms.  Using an EHR and this electronic search algorithm will enable large-scale comparative effectiveness studies on the treatment and management of RA in “real-world” clinical settings.

---

« Back to ACR/ARHP Annual Meeting 2012

ACR Meeting Abstracts - https://acrabstracts.org/abstract/a-validated-mathematical-model-using-electronic-health-records-to-identify-rheumatoid-arthritis-patients-for-observational-studies/