A predictive risk model for first line treatment failure in South Africa

By Julia Rohr  Prudence Ive  Robert C Horsburgh, Rebecca Berhanu, Kate Shearer  Dr Mhairi Maskew  Lawrence Long  Professor Ian Sanne  Matthew Fox  |  | 

Abstract

Background: Although individual predictors of first line antiretroviral therapy (ART) failure have been identified, few studies in resource-­‐limited settings have been large enough for predictive modeling. Understanding the absolute risk of first line failure is useful for patient monitoring and for effectively targeting limited resources for second line ART. The aims of this study are to estimate absolute risk of failure of first line ART over 5 years on treatment as a function of key demographic, clinical, and immunologic factors at the start of ART, and to develop a predictive model that can be applied to other South African clinic populations.  Methods: This is an observational cohort study using medical records from 9 clinics across South Africa, including patients initiated on first line ART after 2004 with at least 6 months of follow-­‐up time. The predictive model for virologic failure on first line (2 consecutive viral load levels >1000 copies/mL) was developed using accelerated failure time models (Weibull distribution), with stepwise selection of potential predictor variables at the start of ART. Multiple imputation was used to impute missing variables. The final predictive model was selected using an internal-­‐external cross validation procedure using Harrell’s C statistic to measure discrimination and difference between 5-­‐year actual and predicted survival to measure calibration.  Results: 71,154 patients were included in the analysis, with an average of 21.5 months (IQR: 8.8-­‐41.5) of follow-­‐up time on first line ART. The final predictive model included age, sex, NNRTI on first line, baseline CD4 count, mean corpuscular volume, hemoglobin, history of tuberculosis, missed visits in the first 6 months on treatment, and an interaction between age and sex. Quintiles of the population were used to create 5 risk groups, where the highest risk group had 24.4% risk of failure over 5 years, and the lowest risk group had 9.4% risk of failure over 5 years. A simplified prognostic score to identify an individual’s risk group was calculated directly from the model parameters.  Conclusions: The predictive model was able to discriminate between patients at higher risk of first line virologic failure. Identification of patients at highest risk of failure is useful for patient monitoring and referral for adherence counseling to improve patient outcomes and avoid the high cost of second line ART.

 

Conference: CROI 2015, Seattle, Washington, USA

Publication details

PDF