The state of the HIV epidemic in rural KwaZulu-Natal, South Africa: a novel application of disease metrics to assess trajectories and highlight areas for intervention

Abstract Background South Africa is at the epicentre of the HIV pandemic, with the world's highest number of new infections and the largest treatment programme. Using metrics proposed by the Joint United Nations Programme on AIDS (UNAIDS), we evaluate progress toward epidemic control and highlight areas for intervention in a hyperendemic South African setting. Methods The Africa Health Research Institute (AHRI) maintains a comprehensive population-based surveillance system in the Hlabisa sub-district of KwaZulu-Natal. Between 2005 and 2017, we tested 39 735 participants (aged 15–49 years) for HIV and followed 22 758 HIV-negative and 13 460 HIV-positive participants to identify new infections and all-cause AIDS-related deaths, respectively. Using these data, we estimated the percentage reduction in incidence, the absolute incidence rate, the incidence-mortality ratio and the incidence-prevalence ratio over place and time. Results We observed a 62% reduction in the number of new infections among men between 2012 and 2017 and a 34% reduction among women between 2014 and 2017. Among men, the incidence-mortality ratio peaked at 4.1 in 2013 and declined to 3.1 in 2017, and among women it fell from a high of 6.4 in 2014 to 4.3 in 2017. Between 2012 and 2017, the female-incidence/male-prevalence ratio declined from 0.24 to 0.13 and the male-incidence/female-prevalence ratio from 0.05 to 0.02. Conclusions Using data from a population-based cohort study, we report impressive progress toward HIV epidemic control in a severely affected South African setting. However, overall progress is off track for 2020 targets set by the UNAIDS. Spatial estimates of the metrics, which demonstrate remarkable heterogeneity over place and time, indicate areas that could benefit from additional or optimized HIV prevention services.

: Shows the (same-sex) female-incidence/female-prevalence and maleincidence/male-prevalence ratios (IPR). Compared with the opposite-sex versions shown in Figure 1, the same-sex incidence-prevalence ratios are less informative about the disproportionate burden of HIV experienced by women relative to men in the study area.
Supplement Tables   Table S1: Summary of the epidemic metrics for all participants (men and women) aged 1549 years in the AHRI surveillance area (20052017 IMR IPR ‡ † N T gives the total number of participants that resided for >50% of the year in the surveillance area (irrespective of consent to HIV testing). HIV+ Prev. and HIV− Prev. represent the HIV-positive and HIV-negative prevalence, respectively. The expected number of HIV-negatives (column 5) is obtained by multiplying N T (column 2) by the HIVnegative prevalence (column 3). The expected number of HIV-positives (column 6) is obtained by multiplying N T (column 2) by the HIV-positive prevalence (column 4). § Shows the number of observed HIV events and person-years of observation (column 7). The HIV incidence rate is per 100 person-years (column 8). The expected number of new HIV infections (column 9) is obtained by multiplying the expected number of HIV-negatives (column 5) by the HIV incidence rate/100 (column 8).
¶ Shows the number of observed deaths among HIV-positive persons and the person-years of observation (column 10). The HIV-related mortality rate is per 100 person-years (column 11). The expected number of HIV-related deaths (column 12) is obtained by multiplying the expected number of HIV-positives (column 6) by the HIV-mortality rate/100 (column 11). The incidence-mortality ratio (IMR, column 13) is obtained by dividing the expected number of new HIV infections (column 9) by the expected number of HIV-related deaths (column 12). ‡ The incidence-prevalence ratio (IPR, column 14) is obtained by taking the average of the male-incidence/femaleprevalence and female-incidence/male-prevalence ratios shown in column 14 of Table 2.

Supplement Methods
This section provides a more detailed overview of our methodology, with slightly dierent mathematical notation than the main text.
Let N T denote the unique number of participants (irrespective of their HIV testing status) that were residents in the surveillance area for more than 50% of the year between 2005 and 2017. Also let i = 1, . . . , N denote the ith participant that consented to an HIV test, such that N ≤ N T . We calculate the HIV-positive prevalence as: where R is the earliest HIV-positive test date, I(R) = 1 if R exists and occurs in year y otherwise I(R) = 0, and N y is the number of participants that tested for HIV in year y. The annual HIVnegative prevalence is therefore H − y = 1 − H + y .
Next, we identied all participants with a rst HIV-negative test followed by at least one HIV test result during the observation period. These repeat-testers comprise the HIV incidence cohort.
To account for the uncertainty of our imputation procedure, we generated [j = 1, . . . , 300] imputed datasets and took the average of the IR y estimates. We obtained standard errors and 95% condence intervals for IR y using Rubin's rules. 3 A well-cited target is to decrease the absolute incidence rate to less than one infection per 1,000 uninfected adults or person-years. 4,5 We calculated the expected number of new infections by multiplying the absolute incidence rate with the expected number of HIV-negative participants in the population: EI y = IR y × (H − y × N T y ). We used this result to calculate the percentage change in the expected number of new infections over a given time-period: where the subscripts y 1 and y 2 denote a baseline year (time 1) and some future year (time 2), respectively. Targets for percentage reductions will vary by country and scale of the local epidemic. For example, under its 90-90-90 treatment targets, the UNAIDS aims to achieve a 75% reduction in the global number of new HIV infections between 2010 and 2020. 6 We next calculated the AIDS-related mortality rate, which is needed for the incidence-mortality metric. Let V ik denote the kth household visit date (k = 1, . . . , K) for participants (i = 1, . . . , N ), where N + is the number of participants that tested Using the same methodology above, we computed geospatial versions of the IR y , MR y , H + y , and N T y . To do this, we used a moving two-dimensional Gaussian kernel of 3 km search radius, 9 the size of which was determined from previous work. 10 We identied the household coordinates of all participants and superimposed their HIV and mortality data on a geographic representation of the study area consisting of a grid of 1 km x 1 km pixels. Next, we calculated Gaussian weighted estimates of the IR y , MR y , H + y , and N T y and generated a raster grid for each. We computed H − y by multiplying the raster grid of 1 − H + y with the raster grid of N T y . Similarly, we computed EI y by multiplying the raster grids of IR y , H − y , and N T . We obtained ED y by multiplying the raster grids of MR y , H + y and N T y . Lastly, we calculated the IMR y for year y by dividing the raster grid generated for EI y by the raster grid generated for ED y , and used a similar procedure for IPR y .