Validation of Algorithms to Identify Human Immunodeficiency Virus Cases Using Administrative Data in Manitoba

Objective: Population-based administrative data are valuable for describing human immunodeficiency virus (HIV) cases, and their health status and outcomes. Our objective was to validate algorithms consisting of physician visits, hospitalizations, and antiretroviral prescriptions against positive confirmatory HIV laboratory tests to identify individuals living with HIV.

Methods: The primary validation cohort consisted of adult Manitoban residents with at least 3 years of health coverage between 2007 and 2018. Positive confirmatory HIV tests from the provincial laboratory were the reference standard. We evaluated 15 algorithms requiring 2 or 3 years of administrative data (hospital, physician, and prescription records) to ascertain cases. Seven measures of accuracy were estimated: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Youden's J, kappa, and area under the receiver operating characteristic curve (AUC) and their 95% confidence intervals. Validity was estimated for pregnant females.

Results: The primary validation cohort included 966,507 individuals, of whom 1452 (0.2%) were HIV cases. Algorithm sensitivity ranged from 82.8% to 97.5%. PPV ranged from 51.8% to 97.8%. Youden's J ranged from 0.83 to 0.97. Kappa ranged from 0.68 to 0.93. AUC ranged from 0.91 to 0.99.

Conclusion: Researchers have a range of algorithms to ascertain HIV cases in administrative data; selection of an appropriate algorithm depends on the user goal. To maximize performance to distinguish HIV cases and non-cases while minimizing data requirements, an algorithm based on three or more physician visits in 2 years is recommended. Further validation in other provinces and territories will assess the generalizability of these findings.