Disentangling the impact of seroconversion age and set-point viral load on ART-free HIV survival

Thesis (Master's)--University of Washington, 2016-06

Introduction: Prediction of off-treatment HIV survival is important in understanding risk among individuals unaware of their HIV status or unable to access care. Historically HIV survival analyses use age at seroconversion or set point viral load (SPVL) as their predictor of interest, but the relationships and interactions between these two covariates has yet to be rigorously determined. We analyzed (1) the impact of different SPVL estimation methods on survival prediction, (2) the relative effects of age at seroconversion and SPVL on survival, and (3) the effect of interaction terms between the two. All models were run on multiple subsets of the same dataset to test for sensitivity to time period, sample size, debiasing methods, and imputation types. Methods: We utilized the CASCADE seroconverters dataset, composed of 16,964 eligible participants. We tested two specifications of SPVL: a geometric mean and a nonlinear modeling method. Our central model was a log-linear regression with time to death (from seroconversion) as the dependent variable and age at seroconversion, SPVL, and an AIDS censorship indicator as the independent variables. We tested five variations on this specification, including a null model, age-only, SPVL-only, a two-way age-SPVL interaction, and a three-way age-SPVL-AIDS indicator interaction. Each of these model specifications was tested on 16 different modifications of the CASCADE dataset: pre-1996 or full-timeseries, debiased or nondebiased, imputed or nonimputed, and testing 18, 20, and 22-year imputation upper bound for the imputed datasets. All models were validated and ranked using 10x10-fold cross-validation with root mean squared error (RMSE) as the error metric. Results: Of the 160 models considered, average RMSE was 3.56 years (range 3.16, 3.98). The nonlinear SPVL method produced estimates with an RMSE 0.29 (0.10, 0.35) years lower than the geometric SPVL method, on average. Models without SPVL performed barely better than the null models. The best-performing model was fit using the nonimputed, nondebiased, pre-1996 dataset, and predicted a 27.0% (95% CI 12.9, 38.8) decrease in survival per tenfold increase in set-point viral load. It did not include a covariate for age. Such a strong impact of SPVL on survival time could have serious implications on mortality for at-risk groups, especially since population-level SPVL has been increasing since the early 1980s. Results were sensitive to the time period modeled. Conclusion: Our analysis showed that SPVL was more predictive of survival than age at seroconversion, but that the way SPVL is calculated has a large impact on predictive performance. We did not find significant effect of the interaction between age at seroconversion and SPVL. Our work highlights the importance of targeting and treating at-risk populations quickly to avoid adverse effects from globally increasing set point viral loads.