Early identification of undiagnosed HIV infection is a critical public health priority. In the United States, nearly 250,000 HIV-infected individuals remain undiagnosed and 50,000 new infections occur annually. Despite several large HIV-related public health initiatives over the last 15 years, the prevalence of undiagnosed HIV infection has remained relatively unchanged. Screening for HIV infection is an important intervention, although controversy still exists as to how it should be optimally implemented. Although the Centers for Disease Control and Prevention (CDC) recommends “routine” nontargeted opt-out HIV screening in most healthcare settings, the United States Preventive Services Task Force (USPSTF) recommends targeted (risk-based) screening. Unfortunately, until recently, little work has been done to determine what characteristics may be used to help identify patients at increased risk of having undiagnosed HIV infection.
Use of both Classification and Regression Tree (CART) and logistic regression analyses are important statistical approaches to empirically deriving models that may help identify patients at risk for HIV infection. This presentation will: (1) introduce the importance of this public health topic; (2) provide a detailed discussion of applied multivariable modeling using both CART and logistic regression modeling for the development of tools to help identify patients with HIV infection using several large datasets from clinical sites throughout the United States; and (3) demonstrate the practical use of these models in clinical care.