Transient increased risk of influenza infection following RSV infection in South Africa: findings from the PHIRST study, South Africa, 2016–2018

We found over twice as many co-infections than expected, if assuming no interaction, and similarly estimate in the model that influenza infection risk is 2.13 (95% CI: 0.97–4.69) higher during or shortly after RSV infection. This, however, translates to a very small population-level impact. In addition, the 95% CIs for the interaction estimate cross 1, so do not meet the traditional threshold of statistical significance. They do however strongly indicate that any interaction is likely to result in an increased infection risk, rather than a competitive interaction as often described in the literature. This result is compounded by very similar output values using the different approaches/methods throughout the paper. Whilst we did not find evidence to suggest elevated RSV infection risk following influenza infection, this may have been due to the small sample size of RSV infections subsequent to influenza. Since the observed synergy in infection risk was estimated to be short-lived, ecological consequences on influenza infection risk following widespread RSV prevention are unlikely; however, they may be amplified by increases in severity as a result of co-infection [2], which we did not address in this paper.

The major strength of this analysis is that the data consist of regular symptom-agnostic swabbing of study participants, allowing the inclusion of mild and asymptomatic infections (which can be a large proportion), rather than just symptomatic and severe cases. This high-quality data results in major benefits. Firstly, changes in infection risk may be obscured by the impacts of dual infection on clinical severity when using syndromic surveillance data, resulting in higher reporting. This has been avoided in our study. The current literature gives mixed results on the increased severity of dual infections, with some studies estimating an increased severity of dual infections [28, 32] and others not finding evidence of a significant effect [33, 34]. The impact of RSV and influenza co-infection on the severity of infection was outside the scope of this study.

The second benefit of symptom-agnostic swabbing is that no bias is introduced as a result of the inherent difference between those with symptoms and those without. For instance, it could be that those with more severe symptoms may have weaker immune systems than those with milder symptoms and therefore may also be more susceptible to infection with both viruses. Including all infections in the analysis removes this potential bias.

Thirdly, by fitting models to infections rather than clinical cases, we reduce the number of parameters required to be fitted, as we do not need to include reporting rate parameters. Fitting complex interaction models is difficult, and it is not possible to differentiate extreme interaction parameter values from less extreme ones [35].

A further strength of the data used in this model is its longitudinal nature. We used data across almost 3 seasons, where the same individuals were sampled within (but not across) each season, giving us detailed infection data. There may, however, be bias introduced due to missing data, with 13% of swabs missing. Whilst the majority of these are likely due to unbiased reasons (such as travel and holidays), other potential reasons, such as severe infections resulting in hospitalisations, may have resulted in bias. However, due to the regularity of the swabbing, any bias introduced due to these reasons would have been minimal, as infection would likely have been identified on a neighbouring swab date (as the median recorded episode duration was 6 and 5.5 days for RSV and influenza, respectively) and very few hospitalisations were found in the cohorts overall. In addition, the hidden Markov model inference takes into account missing data.

Our model results differ from some previous evidence of interaction between RSV and influenza. However, these earlier findings are not based on mechanistic models and could therefore be confounded by other transmission-relevant factors, such as contact rates. This is likely the case with studies investigating causality for the shifts in RSV patterns following the 2009 influenza pandemic [10, 12,13,14], where, despite no governmentally imposed social restrictions, fear of the virus could have altered contact behaviours [36]. In our previous work on evidence of interaction between the two viruses in Nha Trang, Vietnam, we showed that the data was compatible with either a ~ 41% reduction in susceptibility for 10 days following infection with the other virus or no reduction in susceptibility following infection [28]. This is not incompatible with our current estimate of a small increase in susceptibility on the individual level, which would likely not have an impact on the population scale. Our previous model used population-level symptomatic surveillance data, and we did not account for the possibility of cross-protective interaction, as opposed to this study, where the individual-level data allowed us to explore a broader parameter space.

Age-related susceptibility to infection reduction is generally considered to be the case for RSV, whether this be due to age directly or due to subsequent infections. For instance, Henderson et al. showed that attack rates during an epidemic were 98% for the 1st infection, 75% for a second infection, and 65% for a third infection [37]. As increasing numbers of previous infections will strongly correlate with age, this is not out of line with our estimates of age-related hazard rates.

Whilst the evidence we provide shows an increased likelihood of influenza infection following RSV, this is of very small magnitude, most likely resulting in little measurable impact from co-infection at the population level. Of relevance to the implementation of public health policies, a reduction in RSV circulation, for example, due to vaccination, is therefore unlikely to result in substantial replacement with influenza. This provides evidence against a potential concern for the implementation of an RSV vaccine.

Our model was limited by the need to reduce complexity, and we, therefore, did not allow for the waning of immunity, and hence, we were unable to include repeat infection of the same virus within a season or any serological data. Repeat infections of influenza and RSV are likely due to different subtypes. However, as our estimate of interaction is very short-lived, not including these repeat infections is unlikely to have affected our estimates. We were additionally limited by the lack of RSV cases following influenza infections, meaning we had very large confidence in the estimates of the effect of influenza on RSV. We also took a simplistic approach of homogenous population mixing. An alternative explanation of the elevated rate of co-infections is that behavioural aspects may have influenced both the probability of getting RSV and influenza. This could be, for instance, attendance at a crowded event (e.g. religious ceremony) resulting in an increased chance of infection with both viruses. Whilst we may consider that individuals may have higher contacts due to factors such as their employment type, we would expect in this case for the interaction to be longer term, rather than short-lived. We tested this with our ‘high-risk’ sensitivity analysis and found that the short-lived synergistic interaction of RSV infection on the risk of influenza infection was not removed. Ideally, we would be able to quantify times of exposure to the viruses, for example, through infections in the household; however, we did not have sufficient data to analyse this in the current study. Additional differences between individuals could account for some of the enhancing effects that we see, for example, smoking behaviour, indoor air pollution, or inherent susceptibility, we were unable to control for such potentially confounding factors. However, as with different employment types, we would expect these factors to result in a long-term perceived interaction, rather than a short-lived interaction. We also did not look at climatic factors specifically, instead assuming they were captured by our time-varying FOI. This time-varying FOI also allowed us to account for the fact that the influenza season follows the RSV season. We also note that all but one overlapping episode for RSV and influenza were collected during the first season. This is likely due to differing subtypes or age factors, which we were not able to include in our crude analysis of expected dual infections. We attempted to run a sensitivity analysis excluding the first season of data where the majority of co-infections were detected; however, this model did not converge, likely due to small numbers. A further area of exploration would be the impact of rural versus urban environments, yet we also had insufficient data to explore this in the study. We also cannot rule out any residual confounding, so beneficial future work would be to confirm the findings of this study in a different setting.