Background/Purpose: Systemic Lupus Erythematosus is a systemic autoimmune disease that has diverse manifestations that can occur over a long period of time. The complexity of the disease and its varied presentation makes identification of patients with SLE difficult. A better understanding of disease presentation based on clinical aspects of disease could support earlier identification of disease, personalized treatment, improve identification of patients for clinical trials, and support research into the genetic and environmental mechanisms of SLE. EHR data presents a rich source of information, but no detection algorithms have been developed based on “phenotypic” descriptors of SLE. We examined whether the ACR and SLICC classification criteria could be a foundation for the development of such algorithms to detect patients with SLE in EHR data.

Methods: We identified 513 patients with known SLE in a physician validated registry, the Chicago Lupus Database (CLD), whose medical records were also in the Northwestern Medicine Electronic Data Warehouse (NMEDW). We developed algorithms based on ACR and SLICC classification criteria using diagnosis codes (ICD9/ICD10) and lab results to determine whether patients met the same classification criteria in both the CLD and the NMEDW (see Table 1 for criteria). To be classified with SLE, both ACR and SLICC require patients to meet 4 or more criteria, but SLICC also requires 1 clinical and at least 1 immunologic criteria and 4 or more criteria.

Results: As shown in Table 1, of the 513 patients with physician validated SLE present in the CLD who satisfied both the ACR and SLICC classification criteria, the ACR-based EHR algorithm detected 79% (398/513) as having SLE, while the SLICC-based algorithm detected 91% (467/513).

Conclusion: Both ACR- and SLICC-based EHR algorithms detect a significant proportion of patients in the CLD that were classified as having definite SLE. The SLICC-based algorithm had a higher detection rate, likely due to the inclusion of more laboratory parameters in the criteria set and differences in the skin parameter identification, compared to ACR. Our algorithms were developed using only structured data. Both algorithms will likely be improved by using of natural language processing that can probe free-text physician notes for concepts that align with the classification criteria (such as arthritis or renal biopsy results) that were difficult to detect using only diagnosis codes and lab results.