Influenza A (H3) illness and viral aerosol shedding from symptomatic naturally infected and experimentally infected cases

Abstract Background It has long been known that nasal inoculation with influenza A virus produces asymptomatic to febrile infections. Uncertainty persists about whether these infections are sufficiently similar to natural infections for studying human‐to‐human transmission. Methods We compared influenza A viral aerosol shedding from volunteers nasally inoculated with A/Wisconsin/2005 (H3N2) and college community adults naturally infected with influenza A/H3N2 (2012‐2013), selected for influenza‐like illness with objectively measured fever or a positive Quidel QuickVue A&B test. Propensity scores were used to control for differences in symptom presentation observed between experimentally and naturally infected groups. Results Eleven (28%) experimental and 71 (86%) natural cases shed into fine particle aerosols (P < .001). The geometric mean (geometric standard deviation) for viral positive fine aerosol samples from experimental and natural cases was 5.1E + 3 (4.72) and 3.9E + 4 (15.12) RNA copies/half hour, respectively. The 95th percentile shedding rate was 2.4 log10 greater for naturally infected cases (1.4E + 07 vs 7.4E + 04). Certain influenza‐like illness‐related symptoms were associated with viral aerosol shedding. The almost complete lack of symptom severity distributional overlap between groups did not support propensity score–adjusted shedding comparisons. Conclusions Due to selection bias, the natural and experimental infections had limited symptom severity distributional overlap precluding valid, propensity score–adjusted comparison. Relative to the symptomatic naturally infected cases, where high aerosol shedders were found, experimental cases did not produce high aerosol shedders. Studying the frequency of aerosol shedding at the highest observed levels in natural infections without selection on symptoms or fever would support helpful comparisons.


| INTRODUC TI ON
There is uncertainty about the extent to which the site of initial exposure within the pulmonary tree influences influenza virus infection risk and severity between humans. Experimental nasal instillation produces a range of illness from asymptomatic, symptomatic, to febrile. Studies that challenged humans by nasal instillation of virus, and others that challenged with aerosolized virus suggest that upper respiratory mucosal exposure, as opposed to airborne exposure may result in a higher proportion of milder, afebrile illnesses. [1][2][3] Anisotropic infection is defined by Milton as infection whereby transmission mode influences illness presentation, 4 and has been used to characterize human influenza. 5 To minimize health risk associated with experimental human influenza infection, the majority of human challenge models have adopted viral inoculation by nasal instillation. 6 Associations between symptomatology and nasal and throat mucosal viral load following symptom onset have been reported among volunteers receiving intranasal influenza virus challenge and among secondary household cases in Hong Kong. [7][8][9] Other analyses of these household transmission data did not find temporal associations between symptom severity and upper respiratory viral load 10 ; and observed upper respiratory mucosal viral loads 11,12 or respiratory symptoms 13 to be poorly predictive of transmission to household secondary cases, suggesting that other biomarkers of contagion such as exhaled breath aerosols should be explored. The current study compares fine aerosol shedding between influenza A/H3 nasally inoculated and naturally infected cases to test whether experimental, nasal-induced infections have similar risk and rate of fine aerosol shedding compared with natural cases infected by any mode.

| Study design overview
Study design and data collection procedures for the Evaluating Modes of Influenza Transmission (EMIT) human challenge-transmission trial 14 and the observational study of naturally infected influenza cases from University of Maryland campus community are described elsewhere (SI Appendix S1). 15 Half-hour exhaled breath specimens are partitioned into fine (≤5 µm) and coarse (>5 µm) aerosol fractions during collection by Gesundheit-II bioaerosol sampler (G-II). 16 Exhaled breath from both studies was evaluated using standard CDC qRT-PCR primers and probes at the University of

| Symptom scores
Symptom scores were measured three times per day for experimental cases and once per day for natural cases during a research clinic visit where exhaled breath was collected. Scores taken closest in time to exhaled breath collection were selected for analysis.
The upper respiratory score was sum of runny nose, stuffy nose, sneezing, sore throat, and earache symptom scores (range 0-15).
The lower respiratory score was the sum of shortness of breath and cough scores (range 0-6). The systemic symptom score was the sum of malaise, headache, and muscle/join ache scores (range 0-9). The tympanic temperature for experimental and oral for naturally infected cases was recorded. Observed cough counts were recorded during half-hour breath collections.

| Adjustment for qRT-PCR detection limit
Tobit regression was used to impute fine aerosol RNA copy number for qRT-PCR replicates below detection limit where one or more replicates for a sample had detectable RNA. Imputation of RNA copies was not done for samples without any replicates above detection limit, differing from the approach used by Yan and colleagues, where there were a minority of fine aerosol samples below detection limit (14%). 15 It is less reasonable to do the same for the experimentally infected population where 72% of the observations would be imputed. Tobit regression imputed values for samples with qRT-PCR detectable RNA in ≥1 replicate. For both experimentally and naturally infected populations, Tobit models consisted of fixed effects of cough and study day with random in natural infections without selection on symptoms or fever would support helpful comparisons.

K E Y W O R D S
experimental inoculation, human challenge model, influenza symptomatology, influenza transmission, propensity scores, Viral aerosols, viral shedding effect of person. Fixed effects for these models were selected based on a priori evidence of an association with fine aerosol shedding. 15

| Statistics and models to predict shedding
t Tests with equal variances and chi-squared tests were used to compare continuous and categorical demographic and symptom severity variables. Fisher's exact tests were used to compare binomial proportions between experimental and natural cases for fine and coarse shedding subjects and samples. Welsh's t test for unequal variance and the Wilcoxon rank sum test were used to compare geometric mean (GM) and median aerosol shedding, respectively. Tests were two-tailed. Unadjusted effects on probability of shedding into aerosols were estimated with a random effect of person (from generalized linear mixed-effects model) for symptom scores, observed cough count, age, sex, and vaccination status. Analysis of aerosol shedding risk used all exhaled breath observations. Analysis of shedding quantity used the same predictors refined to the maximum shedding day per shedder.

| Case selection and propensity adjustment
Naturally infected cases were sampled from a symptomatic population, selected on the basis of positive QuickVue ® rapid test or febrile illness >37.8°C (measured at University Health Center where some recruitment took place, or upon presentation to research clinic) plus cough or sore throat, and included in analysis based on a single qRT-PCR-positive nasopharyngeal swab on the day of enrollment.
Although enrollment of naturally infected cases used a febrile illness threshold of >37.8°C, this analysis used the more widely accepted threshold of >37.9°C. Experimentally infected cases were selected on qRT-PCR detection of virus from nasopharyngeal swabs on at least two of six follow-up days, or on one day plus serological evidence of infection. Differences in study design were expected to introduce imbalance in symptom severity distribution between groups (SI Appendices S1,S2). If symptoms are associated with aerosol shedding in an unselected population, then this would be an important variable to control for with the goal of assessing differences in shedding between the groups. To minimize the effect of this bias, we attempted to balance covariate distributions between populations with propensity score models (SI Appendix S3).

| Data availability statement
The study data for experimental volunteers are available in the public repository at Nottingham University at https://rdmc.notti ngham.ac.uk/handl e/inter nal/8311 (DOI: http://doi.org/10.17639 / nott.7051). The study data for natural infections are available upon request. All analysis scripts and readme files required to reproduce analyses are available at https://gitlab.com/jacob bueno /natur al_vs_ artif icial_infec tion symptoms score distributions overlapped the most between groups, while differences in the distributions of lower respiratory, systemic, and cough symptoms scores, and cough count were more pronounced ( Figure 2).
For the experimental cases, unadjusted upper and lower respiratory scores, cough symptoms, and cough count were positively associated with fine aerosol shedding detection (Table 3). For naturally infected cases, unadjusted lower respiratory symptom scores and cough symptoms were positively associated with fine aerosol shedding risk. Study day was negatively associated with aerosol shedding in naturally infected cases. Nasopharyngeal swab qRT-PCR cycle threshold (Ct) value had a negative association with aerosol shedding risk for both groups. There were no significant predictors of aerosol shedding rate among experimental or natural cases. Of the experimentally infected, only males shed detectable virus into aerosols. Of the naturally infected, 57/69 (82.6%) unvaccinated and 14/14 (100%) vaccinated with the current influenza season vaccine shed into aerosols above detection limit, while 61/73 with and 10/10 without vaccination for the current and previous influenza seasonal vaccine shed virus into aerosols at measurable levels.
After extensive testing, the best propensity score model (covariates: fever >37.8°C, body temperature, and upper respiratory symptom score) and adjustment by inverse probability weighting for average treatment effect (ATE) minimized the standardized differences between the groups with mean absolute value standardized difference of 91 (Table S5). After ATE adjustment, balance improved for some covariates, however not to the point where they could be considered similar ( Figure 3, Table S5). Although it is advisable that absolute standardized differences for covariates not exceed 10% between comparison groups, and variance ratios approach one (range 0.5-2), 17 the best model with weighted adjustment had absolute standardized differences ranging 7.3%-169.1% and variance ratios up to 48.5. The substantial differences between covariate distributions did not support the use of propensity score-adjusted approaches for making further comparisons between these populations.

| D ISCUSS I ON
We compared aerosol RNA shedding in influenza A (H3) cases infected naturally and by nasal instillation under experimental conditions. Previously, we cultured influenza virus from exhaled breath showing that quantitative culture correlates with RNA copy detection (r = .34, P < .0001). 15 A minority of experimental cases shed virus into aerosols (28%). A far greater proportion of the naturally infected study population shed into aerosols (86%). Among the experimentally infected with detectable viral aerosols, the fine RNA copy GM was within log 10 that for naturally infected cases. A more substantial difference in fine aerosol shedding rate was observed at the level of the overall distribution, with increases in median and upper percentiles for naturally infected cases ( Figure S4). Compared with naturally infected cases at the 95th percentile of fine aerosol shedding, experimentally infected cases shed nearly 2.5 log 10 fewer RNA copies. Given the selection of naturally infected cases on ILI symptoms and/or a positive rapid antigen test, and an observed b Cough symptom score is part of composite lower respiratory score.
c Cough counts after imputation for five experimental and three natural case observations. d Swabs with no detection were coded as having Ct value = 40.
relationship between symptoms and shedding (   There is growing evidence that airborne transmission plays an important role in the spread of influenza. 5,14,18,19 Humans experimentally challenged to influenza virus by airborne particles had a 50% risk of infection to a 0.6-3.5 TCID 50 dose and exhibited increased propensity for moderate to severe illness with fever and cough compared with others experimentally challenged by nasal droplets. 1,2 The term anisotropic has been used to describe such infections where inoculation mode determines illness presentation. 4,5 A population of cases with naturally acquired infections may be ex- unclear that any subsets of cases from each group could be selected to make valid comparisons, we decided to present crude population differences with respect to viral aerosol shedding in Tables 2, 3, and   Table S4, and include everyone meeting case definition (including the 5/39 asymptomatic volunteer cases). We emphasize that all cases had qRT-PCR evidence of infection (Methods).
Identifying naturally infected reference groups that represent the true distribution of symptom severity presents a challenge. Although a substantial proportion of cases are asymptomatic, symptomatic community cases are more prone for inclusion in epidemiologic studies upon seeking medical attention. 20 Multi-year sero-surveil-  10,13,23,24 Clinical detection of infections that may be associated with disease severity and potential for self-isolation or other behaviors that could modify contagiousness should be considered in population-level transmission risk assessment.
Given bias introduced by selection of natural infections on symptoms plus fever or rapid antigen test, the observed correlations of symptoms and fever with viral shedding into aerosols, and a small N with minimal covariable distributional overlap precluding appropriate adjustment, we conclude that the naturally infected population is too different from the experimentally infected cases to make valid comparisons. Our observations show that the 52 nasally inoculated experimental cases produced viral aerosol shedders less frequently than the 83 symptomatic naturally infected population.
When they did shed detectable virus into aerosols, the experimental cases did so at substantially lower quantities than the symptomatic naturally infected group. This difference in aerosol shedding was most pronounced when comparing the highest percentiles of aerosol shedding for each group. The probability and quantity of aerosol shedding in unselected natural infection is unknown. Therefore, these findings encourage efforts to evaluate shedding from infections observed during contact surveillance without selection based on symptoms or fever.

ACK N OWLED G EM ENTS
We

F I G U R E 3
Balance diagnostics for propensity score adjustment by average treatment effect, where the naturally infected cases are "treatment." A, Absolute standardized differences between experimental and naturally infected case covariates plotted for each covariate, the propensity score (PS) and the linear PS. The dotted line represents 10%, which balanced groups do not generally exceed. B, Variance ratios. Dotted lines represent 0.5 and 2, the range for which balanced populations generally do not exceed