Validation of diagnosis codes to identify side of colon in an electronic health record registry.

Abstract

568 Background: The anatomical side of the colon from which a tumor arises is prognostic. Real-world evidence (RWE) studies in colon cancer should include this information. Can International Classification of Diseases (ICD) codes be reliably used to determine side of tumor, and, if so, when? Should electronic health records (EHR) be reviewed to supplement? Methods: Flatiron Health maintains an EHR-derived registry for metastatic colon cancer patients (current N = 9403). A random sample of 100 patients was included in this study. Data about tumor site was compared using ICD codes versus data abstracted from unstructured documents in the EHR (e.g., surgical notes). Concordance was determined via observed agreement and Cohen’s kappa coefficient (𝜿). Accuracy of ICD codes was determined by calculating the sensitivity, specificity, and positive predictive values, and corresponding 95% confidence intervals (CI), using abstracted data as the gold standard. Results: Sampled patients had similar side of colon distribution compared with the full registry: left colon (32% vs. 30%, respectively), right (17% vs. 25%), transverse (7% vs. 5%), unspecified side (39% vs. 37%), and rectum (5% vs. 2%). ICD codes can be general, or specific to the side of tumor. The observed agreement between the ICD codes and abstracted data for tumor site for all sampled patients was 0.57 (𝜿 = 0.42). When restricting to the 56% of patients with a side-specific ICD code, the observed agreement was 0.91 (𝜿 = 0.85). See table for all accuracy estimates. Conclusions: RWE offers growing opportunities for oncology research. Before meaningful RWE analyses can take place, data must be carefully characterized. Here, we describe tumor side EHR data in colon cancer and find that ICD codes specifying tumor side are available over half of the time. When these ICD codes are available, the accuracy of tumor side information in ICD codes is high. When not available, the dataset can be supplemented by chart abstraction. [Table: see text]