Characterizing trachoma elimination using serology

Abstract

Trachoma is targeted for global elimination as a public health problem by 2030. Measurement of IgG antibodies in children is being considered for surveillance and programmatic decision-making. There are currently no programmatic guidelines based on serology, which represents a generalizable problem in seroepidemiology and disease elimination. Here, we collate Chlamydia trachomatis Pgp3 and CT694 IgG measurements from 48 serosurveys across Africa, Latin America, and the Pacific Islands (41,168 children ages 1-5 years) and propose a novel approach to estimate the probability that population C. trachomatis transmission is below or above levels requiring ongoing programmatic action. We determine that trachoma programs could halt control measures with >90% certainty when seroconversion rates (SCRs) are ≤2.2 per 100 person-years. Conversely, SCRs ≥4.5 per 100 person-years correspond with >90% certainty that further control interventions are needed. More extreme SCR thresholds correspond with higher levels of confidence of elimination (lower SCR) or ongoing action needed (higher SCR). This study demonstrates a robust approach for using trachoma serosurveys to guide elimination program decisions.

Fig. 1. Age-specific Pgp3 IgG seroprevalence among 1–9-year-olds.

Evaluation unit (EU)-level seroprevalence to Chlamydia trachomatis Pgp3 antigen among children aged 1–9 years (N = 48 evaluation units, and 63,911 children). Lines represent mean seroprevalence by age estimated using semiparametric cubic splines and EUs are grouped by categories based on programmatic responses (Methods). “Action needed” EUs include populations with clear evidence of ongoing transmission that require public health control measures, while “Action not needed” EUs include populations with demonstrated trachoma control. Unclassified EUs were used as a held-out sample in the analyses. The shaded region in each panel identifies the age range used in the main analyses: 1–5 years (41,168 children). Table 1 includes EU-specific sample sizes.

Fig. 2. Seroconversion rate (SCR) per 100 person-years in 1–5-year-olds.

A Density distributions of the SCR for 34 evaluation units (N = 32,926 children). For each evaluation unit, the black vertical line shows the median estimate, and the density distributions depict the uncertainty about the median. EUs are colored by programmatic response category (Methods) and ordered by increasing median SCR value. The unclassified evaluation units are shown in Supplementary Fig. 1. B Pooled density distributions of the SCR for each category.

Fig. 3. Posterior probability of the need for population-level trachoma interventions using seroconversion rate.

Posterior probability of programmatic ‘Action not needed‘ versus ‘Action needed‘ categories along a range of seroconversion rates (SCRs) among 1–5-year-olds calculated using a two-component Bayesian mixture model (Methods). A Posterior functions assume moderately informative prior probabilities of 80% ‘Action not needed‘ and 20% ‘Action needed‘. In principle, the posterior probability functions allow for the selection of thresholds to inform decisions based on serological surveys with a desired level of certainty. For example, at a ≥ 90% level of certainty, SCR of ≤2.2 per 100 person-years corresponds to a posterior probability of ‘Action not needed‘ and a SCR of ≥4.5 corresponds to a posterior probability of ‘Action needed‘. SCR values > 2.2 and <4.5 per 100 person-years may require additional information to inform programmatic action. B Posterior functions assume an uninformative prior of 50% ‘Action not needed‘ and 50% ‘Action needed‘. Sensitivity analyses in Supplementary Fig. 4 demonstrate that posterior probabilities are insensitive to the prior assumptions.

Fig. 4. Sensitivity analysis of exclusion of evaluation unit- and country-level data.

A jackknife n–1 resampling approach was used to iteratively alter group composition in a Bayesian Mixture model (Methods). A Posterior probability of No Action Needed for trachoma control removing each of 34 evaluation units (EUs) in turn. The red line summarizes the curve fit to all the data, and dark lines show n jackknife subsample fits. In the right panel, red points mark the Seroconversion Rate (SCR) that corresponds with specific posterior probabilities of No Action Needed (in the right panel the Y-axis is zoomed in to the region of 0.5 to 1 to better display estimates). Gray points mark leave-one-out replicates, and open circles with a ‘x‘ symbol indicate the mean SCR over the n jackknife subsamples. Differences between full data estimates (red points) and open circles with a ‘x‘ provide a jackknife estimate of bias, demonstrating no evidence of bias. B Posterior probability curves and SCR estimates that correspond with specific posterior probabilities as in (A), but with all EUs from entire countries left out of each jackknife replicate (N = 10 countries). All estimates assumed an 80% prior probability of no action needed. The two most influential held-out units are labeled in each sensitivity analysis. Overall, there was minimal effect of removing data at EU- or country-level in the higher posterior probabilities (>0.8) – our primary focus. More so, there was an overlap of posterior probabilities and corresponding SCR values of the reduced datasets with that of the original full sample.

Fig. 5. Posterior probability estimates for unclassified evaluation units (EUs).

A Probability of need or no need for trachoma program intervention in unclassified EUs. Unclassified EUs included baseline surveys in new populations that did not have PCR data (Sudan, Peru), opportunistic surveys not focused on trachoma (Malaysia), settings with unusual epidemiology based on trachoma biomarkers (Papua New Guinea, Vanuatu), and those that failed to achieve a consensus classification into ‘Action not needed‘ and ‘Action needed‘ categories (five from Ethiopia and Malawi). The posterior probability was calculated using seroconversion rate (SCR) estimates among 1–5-year-olds in a Bayesian mixture model that assumed prior probabilities of 80% for ‘Action not needed‘ and 20% for ‘Action needed‘. EUs are ordered by increasing median SCR value shown in (B). B EU-specific SCR density distributions, with an example threshold shown at 2.2 per 100 person-years. C An illustrative threshold of 2.2 per 100 person-years corresponding to the 90% posterior probability (’+’ in Fig. 3) was used to calculate the empirical probability of ‘Action not needed‘ as the proportion of the SCR density distribution ≤2.2. Table 1 includes additional details for the unclassified EU populations.