3176 — Evaluating 6 Hepatitis C Case Finding Algorithms using Electronic Administrative Data in the Absence of a Gold-Standard
Taylor TJ, VA Palo Alto Health Care System; Avoundjian T, VA Palo Alto Health Care System; Wu J, VA Palo Alto Health Care System; Asch SM, VA Palo Alto Health Care System, Stanford University; Midboe AM, VA Palo Alto Health Care System;
Structured and unstructured VA electronic administrative data derived from the VA's Corporate Data Warehouse (CDW) are complex and may offer inconsistent reliability and accuracy. Identification of VA patients exposed to hepatitis C Virus (HCV) has become increasingly important as treatment options have improved dramatically. We evaluated 6 algorithms based on previous VA publications for algorithm sensitivity (SE), specificity (SP), negative predictive value (NPV) and positive predictive value (PPV).
VA patients seen in 2014 that had any prior evidence of HCV testing or diagnosis (N = 6,146,987 patients) were assessed via each algorithm within Bayesian 2-latent class analysis. The algorithms were a) comprehensive search across lab testing and ICD-9s (COMP), b) only HCV lab testing, c) expected process of diagnosis/AB testing-RNA testing (EXPECTED), d) 1 inpatient or 2 outpatient ICD-9 (1IN2OUT), e) single ICD-9 diagnosis (1DX), and e) single HCV AB+ test (AB). Diffuse Beta(1,1) priors were used for prevalence estimation and 4 sets of SE, SP, NPV, and PPV priors were evaluated.
Bayesian convergence diagnostics were strong in the selected model (PSRs < 1.02). Specificities were very high ( > .99) for all algorithms within the cohort, though sensitivities varied widely. The order of algorithm performance from strongest to weakest was: COMP (SE = .970, PPV = .992, NPV = .991), LAB (SE = .886, PPV = .991, NPV = .986), AB (SE = .721, PPV = .991, NPV = .975), 1DX (SE = .648, PPV = .991, NPV = .971), 1IN2OUT (SE = .564, PPV = .990, NPV = .965), and EXPECTED (SE = .478, PPV = .990, NPV = .960). Estimated prevalence of HCV exposure was 5.5%, or 383,428 among Veterans active in VA care.
VA CDW-based algorithms may vary greatly in the number of cases identified. Lab and diagnostic processes are highly variable in the case of hepatitis C, and current results suggest a comprehensive search is required to adequately identify patients at risk of HCV related complications.
CDW can be leveraged to identify patients at-risk for clinically and economically complex conditions such as HCV. However, inconsistencies in the data may require comprehensive searches across a wide array of structured and semi-unstructured data elements to achieve high levels of sensitivity; this has implications for resource planning.