Background/Purpose: To harness the
data in electronic health records (EHRs) and administrative databases to study systemic
lupus erythematosus (SLE), it is important to identify patients accurately. Currently,
there is no validated EHR algorithm specifically for SLE. This study sought to develop
and validate a novel EHR algorithm that uses International Classification of
Diseases, Ninth Revision (ICD-9) billing codes, lab values, and medications to identify
SLE patients accurately.

Methods: We used Vanderbilt’s research electronic
database called the Synthetic Derivative (SD).  The SD contains 2.5 million records
with de-identified clinical data from the EHR. There were 5959 potential SLE
cases with at least one SLE ICD-9 code of 710.0.  Of these potential subjects, 200 were randomly selected for chart review to identify the true cases.
A subject was defined as a case if the subject fit the ACR SLE criteria or was diagnosed by a rheumatologist. Positive predictive values
(PPVs) and sensitivity were calculated for combinations of SLE ICD-9 code counts, a positive anti-nuclear antibody (ANA) (titer >
1:40), ever use of antimalarials, corticosteroids, and disease-modifying antirheumatic
drugs (DMARDs), and a keyword of “lupus” in the subjects‘ problem lists. PPVs
were also calculated for excluding ICD-9 codes for
systemic sclerosis (SSc) of 710.1 and dermatomyositis (DM) of 710.3. The algorithm with the highest PPV was internally validated using an additional, randomly selected 100 of the remaining potential 5959 SLE cases. 
Results: The algorithms
with the highest PPVs are shown in Table 1. Excluding the SSc and DM ICD-9
codes improved the PPV as did adding ever antimalarial use, ever DMARD use, and
a positive antinuclear antibody (ANA) to the SLE ICD-9 code. The algorithm with
the best PPV at 95% was 3 or more ICD-9 codes and ANA positive and ever DMARD
and ever steroid use while excluding SSc and DM ICD-9 codes. Using this
algorithm to randomly select 100 subjects for cross-validation, 91 of the 100
subjects were classified as a case resulting in an internally validated PPV of
91%.    Conclusion: We have designed
a novel algorithm that incorporates not only ICD-9 codes but also lab values
and medications to identify SLE patients with a PPV of 91% on an independent test
set.  Being able to identify SLE cases accurately within the EHR will expand
the ability to study more SLE patients from diverse settings. Follow-up studies
are ongoing to externally validate the algorithm in other institutions’ EHRs.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             Table 1.
Positive Predictive Values of Electronic Health Record Algorithms to identify
SLE cases.