Background/Purpose: Pragmatic research in rare diseases is difficult, largely limited by small sample size and single center cohort studies.  The electronic health record (EHR) can serve as a powerful tool to study rare diseases, such as systemic sclerosis (SSc), allowing researchers to collect a large cohort of patients across the healthcare system.  To assemble these patients, algorithms are needed to identify these patients accurately. We have previously developed and validated algorithms to identify SSc subjects in the EHR using ICD-9 billing codes, laboratory data, and keywords.  With the widespread implementation of ICD-10 billing codes, we sought to develop algorithms using ICD-10 codes, as well as other clinical data to identify SSc patients in the EHR. Methods: We analyzed data from a de-identified version of Vanderbilt’s EHR that contains over 2.8 million subjects with longitudinal clinical data.  We identified 1899 potential SSc patients with at least one count of the SSc ICD-9 (710.1) or ICD-10 (M34*) codes.  Of these potential subjects, we randomly selected 200 as a training set for chart review to identify true case status.  A subject was defined as a case if diagnosed with SSc by a Vanderbilt or external rheumatologist, dermatologist, or pulmonologist (specialist).  We selected the following algorithm components based on clinical knowledge and available data: SSc ICD-10 codes, positive anti-nuclear antibody (ANA) (titer ≥ 1:80), and a keyword of Raynaud’s phenomenon (RP) in notes. Positive predictive values (PPVs) and sensitivities were calculated for ICD-10 codes, as well as combinations of the above algorithm components. Subjects with an unclear diagnosis by a specialist (n = 24) or missing clinic notes (n = 20) were excluded from the analysis.   Results:  Table 1 provides the PPVs and sensitivities of the algorithms. PPVs were higher for algorithms using ICD-10 codes compared to those using the ICD-9 code.  Algorithms that used only ICD-10 codes without other clinical data performed well, with PPVs ranging from 94 to 96%.  When adding a positive ANA or RP keyword to ICD-10 codes, PPVs increased to 100%.   The algorithm with the highest combined PPV of 100% and sensitivity of 78% was ≥ 2 counts of the SSc ICD-10 codes and a positive ANA or RP keyword.    Conclusion: Overall, the PPVs for algorithms using ICD-10 codes were higher compared to PPVs for ICD-9 codes.  While adding clinical data to algorithms only using ICD-9 codes was needed to improve those algorithms’ PPVs, this was not needed for algorithms using ICD-10 codes. Therefore, using only ICD-10 codes without the addition of clinical data can serve as an efficient and effective way to identify SSc subjects accurately in the EHR.    

Table 1.

^aICD-10 codes used included M34 grouping.

^b Anti-nuclear antibody (ANA) positive was defined as ≥ 1:80.