Estimating the HIV Incidence Rate – Recent and Future Developments

Abstract

Purpose of review

To describe the needs for information on the number of new HIV infections (incidence) in epidemics and review developments in various methods for its estimation.

Recent findings

Epidemiological methods for estimating incidence with models using prevalence data have been useful but the expansion of antiretroviral treatment programmes could now challenge their reliability. Laboratory-based HIV incidence assays, that can be used to measure HIV incidence using a cross-sectional survey, provide a promising concept but current technologies have not been sufficient accurate. New statistical methods have been developed that show that if the properties of the assay are properly measured then unbiased estimates of incidence can be derived, and that assays meeting certain criteria can produce estimates of acceptable accuracy and precision. Encouragingly, some new assays and algorithms show signs of potentially meeting those criteria. Among the next challenges will be the systematic evaluation assay performance in many different types of specimen and the validation of those methods through comparison with other measurements of incidence.

Summary

Recent developments in epidemiological and incidence assay-based methods of measuring incidence have been substantial and are likely to eventually lead to a revolution in the way that worldwide HIV epidemics are routinely tracked.

Introduction

HIV incidence, the rate of new infections occurring in a population, is the fundamental quantity that specifies the current state of the epidemic. With information about the pattern of incidence in a population, interventions can be targeted to those at the greatest risk of infection, increasing efficiency, while information on trends in incidence over time can be used to evaluate the impact of programmes on the rate of new infections, so that resources can be directed to the most effective interventions. Incidence estimates are also used to the overall trajectory of the epidemic, plan future health care needs and to help design clinical trials.

In contrast to some other diseases, information on the HIV incidence rate is much more valuable than other epidemiological metrics, such as prevalence (the fraction of a population living with infection at a point in time), reported cases or attributed deaths. This is because individuals infected with HIV survive for many years (1, 2) (so changes in prevalence lag changes in incidence) and they may not seek care until many years after infection (so trends in cases reported inform about changes in incidence at prior years (3) (and see (4) and (5) for recent debate on this form of analysis)). Although HIV prevalence is a useful quantity for measuring the overall burden of the epidemic on particular population, it indicates this historic spread of the epidemic rather than the current trajectory and at a population level there can changes in HIV prevalence as the epidemic matures that are not related to recent changes in incidence (6, 7). It has also now been shown that the communities with highest incidence rate do not always have the highest prevalence level (8), meaning that targeting based on prevalence may not lead to the greatest impact of interventions. The same can be true within a population: prevalence among older people is usually lower than those at middle ages, but this can belie a higher incidence rate in older people, and if intervention priorities were determined only on prevalence, then a valuable point of intervention could be missed (9, 10)).

Therefore, it is essential to have information on HIV incidence and recent developments in two of the main approaches – mathematical model and incidence assays – are reviewed here. There are others methods to estimate incidence (including back-calculation using AIDS death, monitoring trends in prevalence among young people or groups of individuals otherwise only exposed to risk for a short time, and trends in case reports) and incidence can be measured in cohort studies, but because this review is focussed on recent development, attention has been focussed on just the two most heavily and researched methods.

Incidence Estimates Using Modelling Based on Prevalence Data

Mathematical models can be used to help untangle the complicated relationship between HIV incidence and prevalence (6, 7). There are two techniques. First, a long-time series of prevalence data can be used to fit a model that relates incidence to prevalence (via assumptions about survival with infection) (11-14). This is what has typically been used to derive the most common estimates of HIV incidence used by UNAIDS in the Epidemic Projection Package (EPP) (15, 16). In this case, the long time-series of prevalence data come from antenatal clinic surveillance (ANC) systems, and this has been supplemented in recent years by estimate of HIV prevalence in the general population obtained in household surveys (16, 17). The advantage of this modelling method is that it can generate estimates of HIV incidence stretching back to the earliest parts of the epidemic, which is particularly useful in estimating the demographic impact of the AIDS epidemic and forecasting future health care needs (18). A limitation is that the model structure must reach a balance between the need to be constrained, so that a plausible incidence trajectory is fitted if there is little data available, but flexible enough that a wide range of possible trajectories can be captured. At first the model used was a simple mechanistic representation of infectious spread (19), but as more data have become available and observed trends have diversified, a model with greater flexibility has been favoured (20). In the next round of estimates, a model with an even ‘lighter touch’ may be used that does not attempt to represent any transmission dynamic and is based only on an spline function that has a large repertoire of possible trends for incidence over time (21).

The same modelling approach has been used to estimate changes in incidence to determine if programmes have had an effect on the natural course of the epidemic (22-24). New analysis and computational techniques have strengthened this approach by allowing uncertainty in derived estimates to be meaningfully quantified (17, 25, 26). However, an important limitation remains that any change in incidence can only be detected in prevalence data retrospectively (22): a reduction in incidence in Zimbabwe in 2000 could only been reliably recorded with prevalence data up to 2007. Looking ahead, this means that, for instance, up to five years after PEPFAR began prevention activities, HIV prevalence data could show little evidence of any impact even if the programmes were truly highly effective (27).

Another limitation of this kind of approach that is now receiving more empirical and theoretical attention is whether it is reasonable to use the ANC data (prevalence among pregnant woman seeking prenatal care) to provide most of the information on the trend in HIV prevalence over time (28-31). This could generate a misleading impression of the trend in prevalence in the general population if any difference between ANC and general population prevalence changes over time, which it might do as epidemics mature (so that more infected women are in late stage infection and so less fertile and underrepresented in the ANC sample) or as epidemics decline (so that the burden of HIV prevalence shifts to older ages under-represented in the ANC sample) or as treatment use expands (and infected women experience increased fertility and enter the ANC sample (32-34) ) (29).

The other set of modelling approaches to estimating incidence from prevalence data have come from Demography (35, 36). Here, serial cross-sectional estimates of prevalence are used to generate an estimate of incidence by building on the principle that a sample of individuals of age a at time t are – under some conditions – representative of a group of individuals age a – τ at time t – τ, where τ is the interval between the surveys. Any difference in prevalence can then be decomposed into the effects of AIDS deaths and new infections. An approximation for AIDS deaths can either be derived with an assumption of a constant age-pattern of incidence or by simply using observed age-specific mortality rates among infected-individuals (37). These methods have been validated by comparing their estimate of incidence with real measurements of incidence in cohort studies (37). They have recently been used to estimate HIV incidence in a number of countries (38, 39) and to determine the age-profile of the burden of infection in projections of epidemic impact (18).

The new challenge for both these modelling approach is the growing influence of ART. The longer mean duration of survival will tend to increase the ‘lead-time’ before changes in incidence are reflected in prevalence (making it harder to interpret prevalence data) but the way in which ART interferes with prevalence trends will actually depend on many factors. For instance, in one survey in South Africa, blood draws were tested for the presence of ARV drugs and 10-20% of the infected population were on treatment, with the highest proportions in those at middle ages and among women (40), but a model suggested that only a fraction (approximately 50%) of those on ART were alive ‘because of ART’ (the others would have been alive anyway) (39). The fraction that would not be alive without ART is influenced by the timing of treatment initiation, the rate of treatment scale-up and mortality rates, which will add substantially to the uncertainty in model-based estimates of incidence, and it should be a priority to validate existing sets of assumptions that account for ART (16, 38).

Incidence Estimates Using Incidence Assays

Estimating incidence using a laboratory assay and data from a single cross-sectional survey has the appealing advantage of immediately providing an ‘up-to-date’ reading on the epidemic. The principle of a test for recent infection is to measure a biological target (“biomarker”) that is related to an early phase of HIV infection (e.g., antibody concentration, proportion, avidity, etc.), and, with knowledge about the time spent in that state, the incidence rate is estimated using the classic “Prevalence = Incidence × Duration” formula. The first such approach was described 12 years ago (41) and progress in the field since then has recently been review by Busch et al. (42) and Mastro et al. (43). Two major challenges are: (i) many such biomarkers are associated with a very large between-individual variation, making it difficult to characterize the mean duration following infection that the marker is detectable, for which accurate and precise knowledge essential; and (ii) that the assays can misclassify a proportion of individuals with long-standing infection as ‘recently’ infected, rather than as chronically infected. The proportion misclassified is termed the ‘False Recent Rate’ (FRR), and if this misclassification is ignored or not correctly factored-in, the resulting HIV incidence estimate can be inaccurate, especially in high prevalence populations, where the number misclassified overwhelms the number of those truly recently infected (44-48). In a recent review of 20 published estimates of HIV incidence using one assay (the BED (49)) that could be directly compared with a ‘reference’ incidence rate, the estimates from the assay were, on average, 27% too high, and, in some extreme cases, more than 400% too high (44). Within-population patterns of assay-derived incidence can also be at odds with other information, such as the older ages of peak incidence recorded in the BED estimates for Uganda (50), when compared to cohort and modelling data (51-53).

In the last year, there has been an evolution in the statistical treatment of data from incidence assays, from approaches that considered the test to be like a diagnostic with ‘sensitivity’ and ‘specificity’ properties for correctly classifying an infection as occuring within a certain period of time after infection (47, 54), to new approaches that model that the existence of the biomarker probabilistically as a function of time since infection. Although a theoretical debate continues about the best way to capture the test characteristics in parameters than can estimated from data (47, 55-61), one method is now recommended by a WHO technical working group (56, 58, 62). This method shows how an unbiased incidence estimate can be obtained (and the uncertainty around that estimate fully quantified), provided there is a local and current estimate of FRR (47, 54, 56, 58). However, because the measurement of the FRR is itself a substantial undertaking (since it requires finding a large and representative sample of HIV-infected individuals known not to have been infected recently), this has not often been done. In a recent review of 39 studies that measured HIV incidence using BED assays, for instance, most did not account for any misclassification (45). Among those studies that did assume a non-zero FRR, the test characteristics used in the calculation were not measured in that population but were instead taken from other studies in different populations (45). That is dangerous because the true FRR varies markedly between populations (e.g. 1.8% in South Africa and 16% in Uganda (44, 46, 48, 63)) and using the ‘wrong’ FRR leads to large errors (46, 47, 64).

The statistical framework also indicates that incidence assays with improved -- but not necessarily perfect -- properties could generate much more reliable estimates. In particular, FRR for a new assay should be confidently measured to be less than 2% for many populations (62). Encouragingly, in early results using specimens from chronically infected individuals in the US, the FRR for one particular type of avidity assay (65) did meet this target, with FRR equal to 1% (66). Further, it has been found that combination of assays used together can have better overall performance and, in preliminary work, the combination of BED and a particular avidity assay (and other virologic and immunological markers) in clade B specimens from the US resulted in an FRR of 0.8% (66). Other new technologies, including cytokine profiles and within-individual viral diversity measures, may be developed into assays over the longer-term. The next task will be to rigorously evaluated and calibrate assays on a range of sample from different clades, populations and epidemic-types and a new specimen repository will be established for this purpose (42, 62).

Next Steps

An important advance in this field would be to develop a statistical framework for combining estimates of incidence from different sources. Currently, this ‘triangulation’ take place informally – if the methods agree, then this increases confidence in the results, but if they do not, then some information is abandoned (63). This seems inefficient because in some ways the different methods’ strengths and limitation are complimentary: for example, ‘level’ information can be reliably extracted from prevalence modelling but that method can be slow to pick up sudden changes in incidence whereas the opposite is the case for imperfect incidence assays, which can fail to accurately estimate the absolute level but which can more readily (and with less calibration required) detect changes over time. AIDS death and case reports and trends in prevalence among different age-groups similarly pull towards the same information from different vantages, and a framework than can unite all these sources, fully representing those strengths, biases and uncertainties, could be very valuable.

Stepping back from the specifics of how to measure incidence, it is worth also considering what the quantities of interest for tracking the epidemic might be in the future. The most important metric might be change in incidence over time or the difference between groups within one population (rather than the absolute level of incidence), whilst the ratio of incidence to prevalence can be a better guides for determining whether the epidemic is under control, and the ratio of new infections to those newly starting treatment might be the more relevant for monitoring national programmes. Whether those indices are ever as useful as incidence is an open question, but it seems likely that identifying and focussing on the most valuable indicators (of which incidence may be only one part, and which may be more attainable than a perfect estimate of absolute incidence) will ultimately lead to better epidemiological intelligence.

Another possibility is the convergence between assays that can be used to estimate the incidence rate at the population level and a potential diagnostic than could indicate whether an individual has been infected recently. This could provide indications on where and among whom new infections are occurring, enhance partner notification and contact tracing interventions (to interrupt infectious spread), help guide the clinical management of patients, support HIV pathogenesis studies, and allow a scientific investigation of the contact network through which HIV is spreading. The technical requirements for different applications are not exactly the same (in particular, the population level assay does not require strong predictive value at the individual level) and progress toward the more difficult target of the individual diagnostic must not hinder the development of a functional population incidence assay. Nevertheless, there would be real advantages of a diagnostic and a substantial market demand would be anticipated for such a tool (52, 54), which could be used to attract the attention of developers and manufactures to the whole area. However, further work is required to: (i) better understand what information about an individual’s timing of infection is generated by assays with different performance characteristics, (ii) the extent to which this would be useful for clinicians and public health interventions.

Conclusion

At this junction in the global pandemic, when efficiency and evaluation of programmes is so critical, familiar tools for measuring incidence have begun to struggle while the new technology of incidence assays is not quite mature. But if, in the coming years, new incidence assay or algorithms can be validated, it could trigger a substantial change in the availability of epidemiological information. The assay may be routinely used in routine household surveys, like the Demographic and Health Surveys, so that a nation’s incidence recent rate is always known. The ‘Know Your Epidemic Know Your Response’ framework that aims to direct programmatic support to those populations most at risk could be updated to allow for real, rather than modelled, values of incidence to be incorporated, allowing a more data-driven and focussed response. (67, 68). It would mean that the impact of national programmes could be easily tracked and it might even facilitate new forms of incentive programme such as Cash-On-Delivery, whereby a payment is made to a country if incidence is reduced (69). And all this would be to the benefit of those that remain at risk of HIV infection.

Key Points.

1. Reliable information on HIV incidence is crucial for an effective and efficient response to the epidemic.

2. Estimates can be derived through modelling prevalence data, but changes in incidence cannot always be quickly detected in this method.

3. Incidence assays hold good potential for measuring current HIV incidence rates quickly and conveniently, but current technologies have not been sufficiently accurate in many settings.

4. Triangulation of different methods may currently be the most reliable approach to estimating incidence.

5. A new generation of incidence assays (or algorithms of assays) may, in the future, provide the best information on HIV incidence rates.