Background/Purpose: Studying births in women with systemic lupus erythematosus (SLE) is difficult given its rarity and the challenges of randomized trials. While the electronic health record (EHR) serves as a powerful tool that is efficient and cost effective, accurately identifying SLE births is challenging.  Our objective was to develop and then externally validate algorithms that use SLE and pregnancy-related ICD-9 and ICD-10 codes, labs, and medications to identify births to SLE patients.

Methods: We used Vanderbilt’s Synthetic Derivative, a de-identified EHR with 2.8 million subjects. We selected individuals with at least 1 count of the SLE ICD-9 code (710.0) or ICD-10 codes (M32.1*, M32.8, M32.9) and at least 1 ICD-9 or ICD-10 code for pregnancy-related diagnoses yielding 433 subjects. For a training set, we randomly selected 100 subjects for chart review. A subject was defined as a case if diagnosed with SLE by a rheumatologist, nephrologist, or dermatologist (specialist) and had a birth documented. Positive predictive values (PPVs) and sensitivity were calculated for combinations of counts of the SLE ICD-9 or ICD-10 codes, ever use of antimalarials, a positive antinuclear antibody (ANA) ≥ 1:160, and ever checked dsDNA or complements (C3 or C4). We performed external validation of ICD-9 algorithms in Duke’s EHR using a set of 100 subjects randomly selected from the Duke Autoimmunity in Pregnancy Registry. ICD-10 algorithms could not be externally validated due to low sample size. Subjects who had an uncertain diagnosis by a specialist (n = 13, Vanderbilt and n = 6, Duke) and subjects with missing notes (n = 5, Vanderbilt and n = 0, Duke) were excluded from the analysis.

Results: PPVs and sensitivities are shown for the algorithms for training and validation sets in Table 1. In the training set, algorithms with ICD-10 codes alone with PPVs from 95 to 100% performed better than algorithms with ICD-9 codes alone with PPVs from 71 to 90%. Adding clinical data improved the PPVs of the algorithms that only used counts of the ICD-9 code. Clinical data only minimally improved the PPVs of algorithms using ICD-10 codes alone. The algorithm with the highest combined PPV of 100% and sensitivity of 95% was ≥ 1 count of the ICD-10 codes and ever dsDNA, C3, or C4 checked. Algorithms using the ICD-9 code and clinical data in the training set replicated well in the external validation set (Table 1).

Conclusion: We have developed and validated algorithms to detect SLE patients with births in the EHR.  Algorithms using ICD-9 codes may require additional clinical data while ICD-10 codes alone can identify SLE patients accurately. Future work is needed to handle subjects with uncertain diagnoses of SLE. Assembling SLE births within the EHR will enable more powerful studies to inform strategies that reduce adverse outcomes.

Table 1.

^aThe training set consisted of 100 subjects from the Vanderbilt EHR.

^bThe validation set consisted of 100 subjects from the Duke EHR.

^cANA positive ≥ 1:160.