Evaluation of ICD codes and phecodes for the identification of pancreatic cancer in a large genomic database.

642 Background: Large genomic databases linked to electronic health records promise to shed light on molecular mechanisms underlying rare diseases, such as pancreatic cancer. However, accurately identifying patients with the desired phenotype can be challenging. This is particularly the case for pancreatic tumors, since ICD codes do not distinguish between pancreatic adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (pNET). Previous studies have shown that ICD codes aggregated by phenotype, known as “phecodes”, have a higher accuracy in identifying specific phenotypes than ICD codes themselves; however, their performance in identifying cancers of the pancreas has not been studied. Methods: From a large deidentified genomic database, two queries were performed to identify all adults with pancreatic cancer for a GWAS study, one using ICD-9/10 codes and the other using phecodes. The medical records for all patients identified from both queries were then reviewed to confirm the presence and histologic type of pancreatic cancer. Results: Of the 91,985 genotyped adults in the database, ICD-9/10 codes identified 1,247 patients with pancreatic cancer, compared with only 422 patients identified by the phecode query. All patients in the phecode cohort were also found in the ICD cohort. Of the 1,247 patients in the ICD cohort, 760 were confirmed to have pancreatic cancer on review of the health records (594 with PDAC, 166 with pNET) whereas in the phecode cohort, only 251 were confirmed to have pancreatic cancer (159 with PDAC, 92 pNET). The positive predictive value (PPV) for PDAC in the ICD query was 47%, compared with 38% for the phecode cohort. The ICD and phecode cohorts had similarly low numbers of pre-malignant cystic tumors (5% in each cohort) and other periampullary cancers (3%). Conclusions: In this large genomic database, the use of ICD-9/10 codes for pancreatic cancer was able to identify nearly three times as many patients with pancreatic cancer and had a higher PPV compared to using phecodes. Therefore, ICD codes, rather than phecodes, should be used to identify patients with pancreatic cancer for subsequent genotyping analysis, though caution is required because the PPV is still low.