Validity of competing algorithms to identify people with HIV in Medicaid administrative claims: a statewide analysis

Administrative claims provide rich data for HIV-related program implementation or research. However, identifying people with HIV in these data remains challenging. We examined the validity of competing case identification algorithms to identify adults with HIV in claims data.

Claims with diagnosis, procedure, and prescription drug codes, as well as enrollment and demographic information, came from Virginia Medicaid, 2012-2023. Surveillance data came from the Virginia Department of Health's Care Markers database, 2012-2023. We created 12 case identification algorithms based on combinations of diagnosis, procedure, and prescription drug codes. After matching the claims to gold-standard surveillance data, for each algorithm, we calculated diagnostic accuracy measures to assess discriminative (sensitivity, specificity, and receiver operating characteristics-area under the curve [ROC-AUC]) and predictive (positive predictive value [PPV] and negative predictive value [NPV]) ability.

For algorithms with a single HIV-related code, algorithm sensitivity was 73-75%, 87% for ROC-AUC, and 47-70% for PPV. As frequency and type of HIV-related codes increased, the sensitivity of algorithms decreased to 58-67% (≥2 HIV-related codes) and 51-63% (≥3 HIV-related codes), whereas ROC-AUC decreased to 79-83% and 75-81%, respectively. PPV increased to 83-88% (≥2 HIV-related codes) and 88-90% (≥3 HIV-related codes). Specificity and NPV exceeded 99% for all case identification algorithms.

Case identification algorithms have good performance when applied in a population-based sample of claims data with low HIV prevalence and using multiple code types. Trade-offs between discriminative and predictive ability suggest algorithms should be tailored to use.