Predicting short-term interruptions of antiretroviral therapy from summary adherence data: Development and test of a probability model

Abstract

Antiretroviral therapy (ART) for HIV is vulnerable to unplanned treatment interruptions–consecutively missed doses over a series of days–which can result in virologic rebound. Yet clinicians lack a simple, valid method for estimating the risk of interruptions. If the likelihood of ART interruption could be derived from a convenient-to-gather summary measure of medication adherence, it might be a valuable tool for both clinical decision-making and research. We constructed an a priori probability model of ART interruption based on average adherence and tested its predictions using data collected on 185 HIV-infected, treatment-naïve individuals over the first 90 days of ART in a prospective cohort study in Mbarara, Uganda. The outcome of interest was the presence or absence of a treatment gap, defined as >72 hours without a dose. Using the pre-determined value of 0.50 probability as the cut point for predicting an interruption, the classification accuracy of the model was 73% (95% CI = 66%– 79%), the specificity was 87% (95% CI = 79%– 93%), and the sensitivity was 59% (95% CI = 48%– 69%). Overall model performance was satisfactory, with an area under the receiver operator characteristic curve (AUROC) of 0.85 (95% CI = 0.80–0.91) and Brier score of 0.20. The study serves as proof-of-concept that the probability model can accurately differentiate patients on the continuum of risk for short-term ART interruptions using a summary measure of adherence. The model may also aid in the design of targeted interventions.