Using serological measures to estimate influenza incidence in the presence of secular trends in exposure and immuno‐modulation of antibody response

Abstract Background Influenza infection is often measured by a fourfold antibody titer increase over an influenza season (ie seroconversion). However, this approach may fail when influenza seasons are less distinct as it does not account for transient effects from recent infections. Here, we present a method to determine seroconversion for non‐paired sera, adjusting for changes in individuals’ antibody titers to influenza due to the transient impact of recent exposures, varied sampling times, and laboratory processes. Methods We applied our method using data for five H3N2 strains collected from 942 individuals, aged 2‐90 years, during the first two study visits of the Fluscape cohort study (2009‐2012) in Guangzhou, China. Results After adjustment, apparent seroconversion rates for non‐circulating strains decreased while we observed a 20% increase in seroconversion rates to recently circulating strains. When examining seroconversion to the most recently circulating strain (A/Brisbane/20/2007) in our study, participants aged under 18, and over 64 had the highest seroconversion rates compared to other age groups. Conclusions Our results highlight the need for improved methods when using antibody titers as an endpoint in settings where there is no clear influenza “off” season. Methods, like those presented here, that use titers from circulating and non‐circulating strains may be key.


| INTRODUC TI ON
Viral detection is the gold standard for measuring incident influenza infections. 1,2 Yet, high asymptomatic infection rates and transient viral presence during infections 3,4 limit detection when using virologic outcomes to measure population-based influenza burden.

Fourfold antibody titer increases over time (ie, seroconversion) and
is traditionally used to measure influenza incidence. 1,2,5 However, an individual's immunological response to influenza combines previous and recent influenza exposures. [6][7][8][9] Current seroconversion methods often do not account for these effects.
Longitudinal sera sampling capture different antibody titer snapshots. Figure 1 illustrates antibody variations before and after infection. Figure 1A shows hypothetical log hemagglutinin inhibition (HI) titers to three influenza strains, sampled antibody response sets over 2 years, and corresponding observed antibody titer changes ( Figure 1B). Ideally, initial sampling captures sera before infection Traditional paired sera (ie, sera sampled at two separate points but tested simultaneously) are ideal, but not always logistically feasible for disease surveillance and longitudinal cohort studies. Using independently tested sera can lead to increased assay-associated error, for example, testing time, protocol adherence, reagent lots, and batch effects, but may be more logistically feasible. 10 Infection also affects the distribution of titer changes, due to possible Collaborator Award (UK, 200187/Z/15/Z to SR), the National Institute for General Medical Sciences (US, MIDAS U01 GM110721-01 to SR), the National Institute for Health Research (UK, for Health Protection Research Unit funding to SR), and the National Institutes of Health Fogarty International Centre (USA, R01 TW008246-01 to SR). strain (A/Brisbane/20/2007) in our study, participants aged under 18, and over 64 had the highest seroconversion rates compared to other age groups.

Conclusions:
Our results highlight the need for improved methods when using antibody titers as an endpoint in settings where there is no clear influenza "off" season.
Methods, like those presented here, that use titers from circulating and non-circulating strains may be key.

K E Y W O R D S
immunodynamics, incidence, influenza, serology F I G U R E 1 Hypothetical changes in antibody titers to three influenza A H3N2 strains at three sampling sets. A, Lines represent hypothetical changes in log hemagglutination inhibition (HI) antibody titers of three infecting strains, where infections occurred at different points during a lifetime. B, shows antibody changes normalized to Strain A (the oldest hypothetical strain to infect an individual). Vertical lines of the same color represent first and second study visit sera sampling points influences like prior exposure or temporary boosting. One probabilistic approach accounts for age to estimate infections, 11 but this approach is computationally intensive and challenging in field settings.
Additional adjustment methods are needed for better across-sample comparisons.
Here, we present an adjustment method using log2 HI titers from multiple recent and historical influenza strains to define seroconversion and measure incident infections. This approach examines the measurement variability in HI serological assays by using the mean titer change. When accounting for an individual's mean titer change, the change to the most recent influenza strain provides a good measure of recent infection while accounting for temporary boosting effects. We apply our method to the FluScape longitudinal study in Guangzhou, China, 12 separately use adjusted incidence estimates to examine demographic effects on the risk of recent influenza infection, and validate our method by applying simulated titer data using a model by Kucharski et al. 11 2 | ME THODS

| Ethics
Johns Hopkins University, University of Florida, and University of Liverpool Institutional Review Boards approved study protocols and materials. Adults, 18 years and older, provided written consent.
Children, 2-17 years, provided verbal assent, and parents or guardians gave written consent.

| Data
Fluscape study participants provided demographic and serological data during the first (December 2009 to January 2011) and second (June 2011 to May 2012) rolling study visits, as previously described. 12 Briefly, eligible individuals were 2 years and older, residing in selected households from 40 randomly sampled communities around Guangzhou, China. Household, individual, and contact questionnaires, and sera were collected at each study visit. Individual questionnaires collected age, gender, occupational status, healthrelated behaviors, influenza vaccination status, and recent influenzalike illness data.

| Laboratory tests
Laboratory methods are previously described. 13 More participants' sera were tested for A/H3N2 compared to A/ H1N1 and B during the two visits; therefore, we used H3N2. We measured duplicate antibody titers using twofold serial dilutions from 1:10 to 1:1280. Positive and negative control sera were also tested.

| Seroconversion
For each individual, we estimated the change in log2 HI antibody titers between baseline and first follow-up visits. Standard seroconversion was a fourfold increase in HI antibody titers (ie, two-unit increase in log2 antibody titers). An individual's estimated adjusted log2 titer change (AC) for each strain was defined as the strainspecific titer change, centered by an individual's mean titer change across all strains: where the titer (T) for i th individuals {i = 1…n} for study visits 1 and 2, j is the influenza A/H3N2 subtype, and K is the number of subtypes tested. We assumed an individual's true baseline titer to older H3N2 strains was less affected by recent infection (ie, is only affected by transient and batch effects), whereas recent infection should increase antibody titers to the most recent strains. We propose that mean-centering individuals' strain-specific antibody titer change across all strains gives a more accurate estimate of changes generated by recent infection. Therefore, mean-centering an individual's strain-specific log2 antibody titer change by their antibody titer change to older strains accounts for temporary boosting and artificial changes from serological testing variations ( Figure 1). Evaluation of two additional methods found no qualitative differences from the proposed method presented here ( Figure S1).
To estimate adjusted seroconversion, we defined a threshold as 2-standard deviations in log2 titer changes from the oldest strain. A/ Hong Kong/1/1968, the reference strain, is the most antigenically distant strain relative to other tested H3N2 strains potentially infecting participants.

| Statistical analysis
Chi-square tests compared baseline characteristics (ie, gender, age, vaccination status, children in household, and residence) by seroconversion. Age group was defined based on influenza risk groups 16 : children (<18 years), adults (18-49 years), older adults (50-64 years), and elderly (≥65 years). Vaccination status was defined as ever receiving an influenza vaccine. Households with children were defined as individuals residing in a household with at least one child (<18 years old).
For households with only one child, that child's household exposure status was defined as not residing with another because they would not be an exposure risk to themselves. Log-odds of seroconversion was modeled as a function of age, gender, vaccination, and children in households using logistic regression. We evaluated the association of age to seroconversion using generalized additive models to estimate spline terms (mgcv package). 19 Age-specific splines used penalized thin-plate regression splines, where estimated degrees of freedom (edf) of knots used penalized likelihood maximization. We also examined the interaction of age groups and presence or absence of children in households on seroconversion. Model fits were assessed using Akaike Information Criteria (AIC). Binomial normal approximation and model coefficient standard errors estimated 95% confidence intervals for seroconversion rates and odds ratios (OR).
We tested our statistical adjustment using simulated data and a published model of influenza antibody titers due to multiple sequential exposures. 11 Six scenarios were simulated using data from Vietnam and China (separate datasets than the one analyzed here).
Simulations included the effects of cross-reactivity from antigenic similarity, long-and short-term antibody boosting generated in previous infections due to subsequent exposure, waning, and antigenic seniority (where previous immunity suppresses responses to subsequent infections). We generated 500 stochastic realizations for 1000 individuals to each scenario and analyzed the data using our adjustment approach described above. Additional simulation details are described in the supplement.

| Sensitivity analysis
We assessed seroconversion rates by vaccination status and seroconversion methods. To identify potential sampling time effects, we examined months between study visits (linear-term) in age-adjusted models of recent infection. Effects of gender and vaccination status on recent infection risk were also evaluated.

| Study participant characteristics
During two study visits, 2012 participants from 856 households provided household and individual demographic data. 1018 participants had paired sera available for antibody testing. Previous comparisons found no demographic differences between those who did and did not provide sera. 12 Of those with paired sera, 942 participants had sera available for all five A/H3N2 strains for both visits (

| Strain-specific antibody titer changes
Overall, when examining mean titer changes across the five strains, the distribution from the standard method was left-skewed and

| Effects of individual and household factors on the risk of recent infection
We

| Method validation
Our method validation used simulated data from 500 stochastic real-  (Table S6). In the presence of antigenic seniority, both methods underestimated the true infection status, but the adjusted method estimated seroconversion rates 1.5 to 5.6 times higher than standard rates depending on the infection probability, more closely reflecting true infection rates compared to the standard method. ness, given HA group imprinting can provide cross-immunity. 9,29 Our methods rely upon having results from multiple strains. We expect our results to work better when more strains are included and expect that our methods may fail if only a small number of strains are used. Accounting for past exposures will be critical in future evaluation of seroconversion, and our method may also apply to other viral families with cross-reactivity. 30,31 Our findings show our proposed method for measuring incident influenza infections improves seroconversion estimates to recently circulating influenza strains and in validation, estimated seroconversion rates closer to the true infection rate in the presence of antigenic seniority. Our approach may have relevance for assessing incidence in cohort or sero-surveillance studies, when testing paired sera maybe logistically challenging. We highlight the need to consider effects of multiple viruses on antibody responses over the lifecourse. Since strains circulating earlier in an individual's life continue to influence responses to recent and antigenically-related subtypes, examining antibody titers to one strain using standard seroconversion methods may fail to fully capture true incident events.

CO N FLI C T O F I NTE R E S T
The authors declare that they have no conflicts.