Candidate for the Lee B. Lusted Student Prize Competition
Purpose: International guidelines recommend early HIV treatment initiation (i.e., at CD4 <350) for HIV-infected individuals in resource-limited settings. However, funding availability for early or deferred HIV treatment (i.e., at CD4 <200) in Haiti is uncertain. We aimed to internally validate and calibrate a user-friendly model of HIV disease in Haiti that will assist policy makers in forecasting treatment need and capacity.
Methods: We used patient-level data from Haitian observational cohorts and a randomized trial conducted in Haiti to develop a computer-based, mathematical model of HIV disease. Incidence density analysis was used to derive model parameters for untreated HIV disease progression (HIV seroconverters cohort, n=41; asymptomatic HIV disease, n=436) and HIV treatment (early 1st-line treatment, n=408, deferred 1st-line treatment, n=910; deferred 2nd-line treatment, n=194). Model predictions were compared to observed data to assess internal validity. Goodness of fit measures included visual assessment of Kaplan-Meier survival curves, comparisons of 5-year event probability, and percentage deviation between the predicted estimates and observed data at discrete time points, averaged over time. When model predictions did not exhibit a good fit due to model structure simplifications that would enhance usability, an internal calibration algorithm was applied to improve goodness of fit between predicted and observed outcomes. The model was implemented in Microsoft Excel, and results evaluated over a 5-and 10-year policy time horizon.
Results: For a cohort of newly HIV-infected individuals with no access to HIV treatment, the model predicts median AIDS-free survival of 9.0 years pre-calibration and 5.6 years post-calibration versus 5.8 years (95% CI 5.1, 7.0) observed (Figure 1). For a cohort of patients initiating deferred treatment, the model estimates 23.2% would die by 5 years (versus 23.5% in the observed data), 7.3% would be lost from care (versus 7.8%), and 11.7% would initiate a second treatment regimen (versus 10.8%). In 12 out of 14 comparisons assessing different natural history and treatment-related outcomes, mean percentage deviation between the model predictions and observed data does not exceed 5% over both 5 and 10 years.
Conclusions: Internal validation and calibration results were sufficient for 5- and10-year health policy decision making. Using local data in a model-building process can improve validity and acceptability of policy models in resource-limited settings.