Estimating HIV, HCV and HSV2 incidence from emergency department serosurvey

Data

At the adult JHH ED, identity-unlinked sero-surveys were conducted during six to eight-week periods in 2003, 2007, 2013 and 2016. These surveys are described in detail elsewhere^21,25. Briefly, excess sera were collected and assigned a unique study ID while chart review data were recorded in real-time. All laboratory testing was done after the collection period when the linked patient identifiers were removed from the dataset. Patient consent was waived by the ethical review board. This study was approved by the Johns Hopkins School of Medicine Institutional Review Board (IRB00083646, CIR00016268) and conducted by the ethical standards of the Helsinki Declaration of the World Medical Association. In the current analysis the authors had access only to anonymized data consisting of the test outcomes, sex, race (coded as black, white or other), age category (yearly from 18 to 89, and over 89) alongside the year the test was performed. The analysis excludes multiple results per unique subject, those < 18 years of age and individuals that did not self-identify as black or white, male or female. The reason for this limitation is due to sample size limitations. We also excluded 31 patients with incomplete data for one or more tests. Prevalence summaries of the data, showing the changing age-profile of infection, are provided in Figure 1, and aggregated data is shown in Table 2.

Figure 1. Prevalence of infection by disease, race, gender, age-group and survey year.

Demographic data / death rates

In order to accurately model the movement of individuals through the various infection (and co-infection) classes, we need to know the excess mortality due to infection with HCV or HIV. HIV-specific mortality rates are taken from 26, a study estimating the age-stratified mortality of HIV positive individuals on antiretroviral therapy (ART). These rates may underestimate the excess mortality in our population since: the rates are not race-stratified; and the study population are all on ART. Conversely, these rates may overestimate mortality since: individuals on ART are unlikely to be recent infections; the study includes individuals with HCV; and the rates are not described relative to the background death rate. Table 1 gives the age- and year-specific mortality rates taken from 26. We assumed that the mortality rates remained constant after 2008.

HCV-specific mortality rates were obtained from Mahajan et al., a cohort-study of HCV- infected patients which reports an annual mortality rate of 12.854% in the cohort, compared with 1.046% in the general population²⁷. However to be included in the cohort, patients must be aware of their HCV status, which is not the case for the JHH ER data. Denniston et al. report that 50.3% of people detected with HCV were unaware they were infected²⁸. We assumed that infected individuals needed to be aware to be included in the cohort described in Mahajan et al. and, since patients that are aware are more likely to have experienced symptoms, we assume that unaware individuals experienced a death rate similar to the general population. This results in a HCV mortality rate of 5.868% annual excess death rate. Finally Thomas et al. find that end-stage liver disease is 3.67 times more likely in those over 38 years of age²⁹, and so we reduce the death rate from 5.868% for over 37s to 1.6% for under 38s.

Model

We developed a novel cohort model (illustrated in Figure 2), in which individuals are born with no infection, and then, over their lifetime may acquire each of the three infections, in any order. Each vertex of the cube in Figure 2 represents a state an individual may be in. Each state is denoted by three digits, giving the status of each different disease (0 if uninfected with that disease and 1 if infected). The digits are given in the order: HCV, HIV and HSV2 so that, for example, 101 is the state of being infected with HCV and HSV2 and uninfected with HIV. We use this notation to write down a system of eight ordinary differential equations (ODEs), tracking the proportion of the population in each state over time. We divide the population into “cohorts” of individuals of similar ages and for each cohort we use the same eight ODEs, but allow the parameters (the λ’s and θ’s) to be different for different cohorts. Each disease has a baseline infection rate for uninfected individuals (λ₁, λ₂ and λ₃). The increased (or decreased) risk of infection due to previous infection status is given by the θ’s. For example, the relative risk of infection with HIV of an individual that is already infected with HCV (but not HSV2) is given by θ₁. If θ₁ is greater than (less than) 1, we expect that prior infection with HCV increases (decreases) the risk of infection with HIV.

Figure 2. Model schematic.

The schematic shows individuals moving between uninfected (0) and infected (1) for each of the infections: HCV; HIV; and HSV2, in that order. The rates of infection are represented as a varying force of infection, λ's for each disease. Where individuals are already infected with at least one of the viruses, these forces of infection are modified by a factor, θ, which could increase or decrease risk.

In addition to the events shown as arrows in Figure 2, we also include a variable death rate due to disease status. Since the natural death rate affects all disease states equally, without changing the proportions of the cohort in each state, we only need to consider the increase (or decrease) in death rate due to a particular disease status. For example, if infection with HIV increases the rate of death, then we would expect the proportions of the cohort in states 010, 110, 011 and 111 to decrease over time relative to the other states due to this.

We take p^ijk to be the proportion of the total population that have status i for HCV, j for HIV, and k for HSV2. So, for example, p⁰¹¹ is the proportion of the population that are negative for HCV and positive for both HIV and HSV2. Then our model is given by the following ordinary differential equations (ODEs):

where $p_{\text{HCV}}={\displaystyle {\sum }_{i=0}^1{\displaystyle {\sum }_{j=0}^1{p}^{1ij}}}$ and $p_{\text{HIV}}={\displaystyle {\sum }_{i=0}^1{\displaystyle {\sum }_{j=0}^1{p}^{i1j}}}$. Here δ_HCV and δ_HIV represent the additional risk of death due to infection with HCV and HIV, respectively.

Note that since we are considering proportions of the total population, the primary way that the additional risk of death is seen in the equations is through the reduction of total population size. This corresponds to an increase in each of the other proportions due to the reduced denominator.

Bayesian inference approach

The infection rates (incidences) were allowed to vary smoothly with time and age cohort. This flexible approach allows trends in incidence to be inferred from the data. The parameters are estimated in a Bayesian framework using an adaptive Markov chain Monte Carlo algorithm³⁰ and implemented in R³¹. The prior for the incidence parameters is a first order Gaussian random walk³², which implies that the difference between incidences at consecutive timepoints/cohorts is Normally distributed with mean zero. This is analogous to a Gaussian process³³ evaluated at discrete locations. The variances of the normal distributions describe the smoothness of the incidence surface, and are also estimated from the data. There are six smoothness parameters for each gender and race cohort, one for time and one for age for each of the three viruses. The nine relative risk parameters (the θ’s, see Figure 2) and the initial state of the model in 2003 were also estimated from the data.

We model the proportion of individuals with each infection status separately for each age cohort: birth year before 1943, and then each subsequent birth year up to 1995. We take a Bayesian non-parameteric approach and estimate the three infection rates (incidences) for each year of the study (2003 to 2013) and for each age cohort. We therefore produce three incidence surfaces across time and age cohort. Because there is only a small amount of data for each cohort, we assume that the incidence surfaces are smooth, and so they can “borrow strength” from neighbouring points. Although estimating these incidence surfaces is the main challenge of the inference, in addition we must also estimate the 9 relative risk parameters (θ₁, . . . θ₉), the initial conditions for each age cohort and the hyperparameters that describe the smoothness of the incidence surfaces.

More precisely, let λ_i,j,k represent the rate of infection for disease i in year j for age cohort k, for i ∈ {1, 2, 3}, k ∈ K and j ∈ {2003,..., 2015} ∩ A_k, where K is the index set of age-cohorts and A_k is the set of study years after which cohort k reaches 18 years-old. The infection rates can be interpreted as the proportion of a completely susceptible population that would be infected in a year, and therefore we refer to them as incidences. Since we have data at the yearly resolution, we assume that the incidence is constant within each year.

Let ${\mathit{p}}_k=\left(p_k^{000},p_k^{001},\mathrm{...},p_k^{111}\right)$ represent the vector of initial conditions for age-cohort k, which must sum to 1. These vectors represent either the disease status at the start of the study in 2003; or, for the younger cohorts, the disease status when the cohort turns 18.

Priors. We assume a first order Gaussian random walk prior for the incidences across both time and age-cohort. For all i ∈ {1, 2, 3},

This definition includes 6 hyperparameters (κ’s) that control the smoothness of the surfaces. We assumed that independent Gamma distributed priors with shape parameter 1 and rate parameter 0.01. Since this Gaussian random walk prior only specifies the differences between the lambdas, we also assumed that the mean level of each incidence surface follows an exponential distribution with mean 1%, to provide a mild shrinkage effect on the mean incidences.

The priors for the initial conditions were assumed to be Dirichlet distributions (which ensures that they sum to 1) with parameters equal to α = 0.8/8, following the uninformative prior suggested by Berger et al.³⁴. The priors for the relative risk parameters θ₁, . . . θ₉ were assumed to be independent exponential distributions with mean 1.

Likelihood. We assumed a multinomial likelihood function. Let $d_{j,k}^m$ be the number of individuals that tests indicate are in disease state m ∈ {(000), (001), . . . , (111)} in year j from age-cohort k. Let d_j,k represent the data vector for cohort k in year j and θ represent the complete vector of parameters. To evaluate the likelihood for cohort k in year j we start with the initial conditions p_k and solve the differential Equation (1)–Equation (8) forward in time until year j, call this p_k(j). This was done using the rk4 Runge-Kutta differential equation solver from the R package deSolve³⁵. The log likelihood of observing data d_j,k given p_k(j) is simply ${\displaystyle {\sum }_m{d}_{j,k}^m\log \left(p_k^m\left(j\right)\right)}$.

The complete log likelihood is then just the sum over all these component parts.

Markov chain Monte Carlo algorithm. To draw samples from the posterior distribution of the model we used Markov chain Monte Carlo. To update the λ_i,j,k we used two kinds of single-site proposal. For j and k not on the boundaries we used conditional prior proposals³⁶ for 50% of the updates. For all remaining updates we used a Metropolis-Hastings Gaussian random walk, with proposal variance automatically tuned to 44% by rescaling by a factor x_n at iteration n when the iteration is accepted, and by $x_n^{\text{0.44/(0.44−1)}}$ when iteration n is rejected. To achieve diminishing adaptation, we used the sequence x_n = 1 + 20/(20 + 0.2n). The precision parameters controlling the smoothness were updated using a Gibbs step.

If we consider only the data for the initial time point, the vector of initial conditions for cohort k, p_k, has full conditional distribution p_k|d_1,k ~ Dir(α1 + d_1,k). For the final cohort there is no subsequent data and so the full conditional can be used as a Gibbs step. For earlier cohorts we cannot evaluate the posterior without first solving the system of differential equations. To find an efficient proposal we developed our own adaptive algorithm for Dirichlet distributions (similar to the algorithm in 37) that balances the above conditional distribution against the current location of the chain. Our proposal for p_k was ${\mathit{p}}_k^{\prime }$ ~ Dir(α1 + d_1,k + β_kp_k), which was then accepted or rejected based on the usual Metropolis-Hastings ratio. To find an appropriate value for β_k for each cohort k, we applied the following adaptation algorithm during the burn-in phase of the MCMC. If a proposal was accepted then β_k ↦ max{0, β_k − 3}, and if a proposal was rejected then β_k ↦ β_k + 1. This leads to an acceptance rate for p_k of roughly 25%, unless there is little data after the initial condition, in which case β_k ≈ 0 and the acceptance rate is much higher.

The relative risk parameters θ₁,..., θ₉ were updated with single site Gaussian random walk proposals. To improve the mixing between these highly correlated parameters, we also added a joint update of the θ’s and the means of the 3 incidence surfaces using a Metropolis-Hastings Gaussian random walk in 12 dimensions, with the proposal covariance estimated adaptively using the approach of 30, implemented using the accelerated shaping algorithm³⁸.

Incidence

Figure 3 shows the posterior median incidence rates in uninfected individuals for the three diseases as a function of time and age-cohort, with lower and upper confidence intervals shown in Figures S2, S3 and S4 (Extended data³⁹). For most groups and most infections we see a much stronger age-cohort effect than a time effect. That is, incidence in a cohort of individuals born in the same year remains approximately constant over time, while an overall decrease in incidence results from lower incidence in younger age groups. This effect is particularly prominent in HCV in black females, black males and white males and in HIV infection in black females (Figure 3). Incidence of all three infections is generally higher in the black population (the first and third rows in Figure 3) than the white population (rows two and four). In most graphs we see a small increase in incidence in the youngest age group in the most recent year. However, this is likely to be a statistical artefact resulting from increasing uncertainty due to fewer age groups (and thus less data) in more recent years. This increase in uncertainty when incidence rates are low can lead to an increase in the median, and so the uncertainty in the estimates is important in these regions (see Figures S1, S2 and S3 of the Extended data for confidence intervals³⁹).

Figure 3. Incidence rates.

Inferred incidence rates of Hepatitis C Virus (HCV), Human Immunodeficiency Virus (HIV) and Herpes Simplex Virus (HSV2) infection amongst black females (BF), white females (WF), black males (BM) and white males (WM) as a function of time and year of birth.

The first column of Figure 3 shows HCV incidence. Black females, black males and white males all display a strong cohort effect, with individuals born around 1960 experiencing consistently high incidence, while younger cohorts show progressively lower incidence. The posterior median annual incidence peaks at 3.22% (90% CI: 1.66 – 4.76%) in 2003 for black males born in 1953. The peak incidence for black females was 2.95% (1.45 – 4.42%) for those born in 1957 and 2.63% (1.35 – 4.06%) for white males born in 1961, both also in the earliest year in the study. In contrast white females have consistently low incidence, decreasing in all cohorts to between 1.09 and 1.87% for the most recent year.

The second column of Figure 3 shows HIV incidence. HIV incidence is more constant than HCV, and is not as smooth. Incidence in black females and white females is quite constant over time with posterior median incidences in the ranges 1.04–1.91% and 0.52–1.59%, respectively. However, the younger cohorts of black females appear to display somewhat reduced incidence. There is also some indication of reduced incidence in black males in both younger cohorts and more recent years. For black males born in 1962 the posterior median incidence drops from a peak of 1.94% at the start of the study to 1.40% in 2012; however, these declines are small compared to the level of noise (see Figure S2 of the Extended data for HIV confidence intervals³⁹).

The third column of Figure 3 shows posterior median HSV2 incidence. Black females show increasing incidence of HSV2 for later birth cohorts, peaking at 8.34% (90% CI: 5.09–11.5%) in the latest cohort, although there is some suggestion of decreasing risk over time for each cohort. This indicates a higher risk for younger individuals, but decreasing slightly as they age. In contrast, white females’, black males’ and white males’ incidence of HSV2 does not change as much as for black females. White females also show increasing risk for later born cohorts, with a maximum of 3.82% (2.41–5.16%) for individuals born in 1978 and then a decrease for the most recent cohorts. For each cohort, incidence decreases somewhat over time. A similar effect is seen in black males, with the highest-risk cohort born in 1973 with an incidence of 3.61% (1.99–5.26%) and general decreasing incidence over time for each cohort. Black males also display a low-risk cohort born around 1960, with incidence dropping to 1.81% (0.51–3.34%) for those born in 1958. Incidence in white males again shows decreasing risk over time for each cohort and relatively flat incidence between the cohorts.

Relative risks

Table 3 shows the posterior values of the θ parameters (see Figure 2). These give the relative risk of acquiring an infection for individuals that already have another infection. For example, θ₁ gives the relative risk of acquiring HIV for individuals that already have HCV. If θ₁ > 1 (or θ₁ < 1) then this would indicate that individuals with HCV are more likely (less likely, respectively) to acquire HIV than uninfected individuals.

Individuals that already have HSV2 are at lower risk of acquiring HIV or HCV (θ₅ and θ₆ < 1). This has the possible explanation that individuals that are infected with HSV2 before any other disease are not therefore at high risk of acquiring HIV or HCV. Conversely, individuals that acquire HIV or HCV first are at greater risk of becoming infected with HSV2 (θ₂ and θ₄ > 1). For those individuals eventually coinfected with HIV and HSV2, if they are infected with HIV first, they rapidly become infected with HSV2. The parameter θ₇ represents a similar mechanism, however due to infrequent occurrence of coinfection with both HIV and HCV (due to high death rates), there is insufficient evidence to determine whether θ₇ is greater than or less than 1.

The other θ parameters are mostly inconclusive, but there is a suggestion of reduced infection with HIV when already infected with HCV (θ₁ < 1), reduced infection with HIV when already infected with both HCV and HSV2 (θ₈ < 1) and increased infection with HCV when already infected with both HIV and HSV2 (θ₉ > 1).

Forward projections

We extended the random walk prior on the incidences to predict the incidences from 2014 to 2016, with associated uncertainty. The model equations were then solved from 2013 onwards to produce the posterior predictive distribution of prevalences in 2016 with the data from 2016 held-out for validation purposes.

Projections for 2016 were generally quite good for HCV and HIV (bottom rows of Figure 4, Figure 5, Figure 6 and Figure 7, whilst in general overestimating for HSV2. For white males and females the HIV prevalence was very low and the prevalence was overestimated in 2016.

Figure 4. Model fit and predictions for black females.

Posterior median prevalence for black females (solid line) and 90% credible intervals (shaded) line for three diseases (columns) in the four sero-surveys (rows), each separated by age cohort. The model has been fitted to the observed data (crosses) in the first three rows of plots but data in the bottom row was held-out for validation purposes. Vertical lines represent 90% binomial CI for the observed data.

Figure 5. Model fit and predictions for white females.

Posterior median prevalence for white females (solid line) and 90% credible intervals (shaded) line for three diseases (columns) in the four sero-surveys (rows), each separated by age cohort. The model has been fitted to the observed data (crosses) in the first three rows of plots but data in the bottom row was held-out for validation purposes. Vertical lines represent 90% binomial CI for the observed data.

Figure 6. Model fit and predictions for black males.

Posterior median prevalence for black males (solid line) and 90% credible intervals (shaded) line for three diseases (columns) in the four sero-surveys (rows), each separated by age cohort. The model has been fitted to the observed data (crosses) in the first three rows of plots but data in the bottom row was held-out for validation purposes. Vertical lines represent 90% binomial CI for the observed data.

Figure 7. Model fit and predictions for white males.

Posterior median prevalence for white males (solid line) and 90% credible intervals (shaded) line for three diseases (columns) in the four sero-surveys (rows), each separated by age cohort. The model has been fitted to the observed data (crosses) in the first three rows of plots but data in the bottom row was held-out for validation purposes. Vertical lines represent 90% binomial CI for the observed data.

Strengths and limitations

The main strength of our approach was that we were able to infer cohort-specific incidence rates that varied through time, and therefore were able to identify trends in incidence at the age, gender and race cohort level. Our results showed that there could be large differences in incidences between age cohorts, which may not have been noticeable from a more naive analysis of the data.

A key weakness of our method was that for our incidence rates to be well-calibrated, we require an accurate estimate of the death rates in HIV and HCV infected populations, over and above the death rate in uninfected populations. This is because the death rates and infection rates are not both identifiable from the proportion of individuals in each infection state. The death rates proved challenging to obtain from the literature, particularly as HIV death rates may have improved substantially in recent years. Nonetheless, if these death rates are stable but incorrectly estimated, any qualitative trends in incidence informed from the data will remain correct.

Underlying data

Zenodo: drsimonspencer/HIV-HCV-HSV2-coinfection: Source code and extended data. https: //doi.org/10.5281/zenodo.5078271³⁹.

This project contains the following underlying data:

Extended data

Zenodo: drsimonspencer/HIV-HCV-HSV2-coinfection: Source code and extended data. https: //doi.org/10.5281/zenodo.5078271³⁹.

This project contains the following extended data:

Data are available under the terms of the Creative Commons Zero ”No rights reserved” data waiver (CC0 1.0 Public domain dedication).