Assessing the use of surveillance data to estimate reductions in HIV incidence achieved by combination HIV prevention programs: a mathematical modelling study


BACKGROUND: Surveillance data on new HIV diagnoses are frequently used as a proxy for HIV incidence when assessing HIV intervention programs. We used mathematical modelling to determine when changes in diagnoses or other surveillance measures could reliably approximate HIV incidence changes for evaluation of combination HIV prevention programs.
METHODS: We used a calibrated model of HIV transmission, antiretroviral therapy (ART) and pre-exposure prophylaxis (PrEP) among men who have sex with men in Baltimore, US, to simulate ten-year combination prevention programs expanding ART, PrEP and HIV testing together, with a one-year scaleup period. We determined how well modelled relative changes in annual HIV incidence (compared with pre-program levels) could be reflected by relative changes in total annual HIV diagnoses or other surveillance measures ' diagnoses with acute infection, diagnoses adjusted for testing volume, the proportion virally non-suppressed - at different timepoints. We report the median (95% credible interval) absolute bias of each measure in percentage points (pp).
RESULTS: Modelled changes in total HIV diagnoses underestimated declines in new HIV infections, by -25pp (with substantial variability: -116,-5pp) in the second year of the prevention program, 3pp (-17,-1) in year 5, and -1pp (-5,+1) in year 10. The extent of the bias was positively correlated with the increase in levels of diagnosis achieved by the program. Declines in diagnoses with acute infection always underestimated declines in incidence (by -27pp in year 2 and -10pp in year 10) with considerable variability. Adjusting diagnoses by test volume somewhat reduced biases in year 2 although they remained variable [bias +9pp (-6,+24)], and overestimated incidence declines in year 10 [by +14pp (+3,+34)]. Changes in the proportion virally non-suppressed gave unbiased but variable estimates of incidence declines in year 2 [-2pp (-19,+11)], but underestimated incidence declines in year 10 [-10pp (-31,+0)].
CONCLUSIONS: When evaluating combination HIV prevention programs which expand HIV testing, changes in annual total diagnoses do not reflect incidence reductions well until several years into the program. Changes in diagnoses adjusted for testing volume or the proportion virally unsuppressed can give less biased (although still variable) estimates of incidence changes earlier on in the program.