The average age of infection is expected to vary during seasonal epidemics in a manner dependent on important epidemiological features, such as the duration of infectiousness and the nature of transmission-relevant mixing. Age–incidence patterns result from fluctuations in the susceptible population that vary by age combined with age-dependencies in the transmission rate, both of which influence what age group is responsible for initiating seasonal epidemics (see electronic supplementary material) [4
]. We have shown here that seasonal changes in the average age of cases could be detected using routinely collected data, and were consistent with model predictions. Examining age–incidence patterns can lend insight into important epidemiological features of infections.
We found that simple models for the transmission dynamics of measles and rotavirus could predict the lag time associated with the maximum correlation with a margin of error of ±8–12 weeks. Improving the models by incorporating more information about the epidemiology of infection led to a better correspondence between models and data, such that predicted lag times were within ±1 week of those observed. The amplitude of seasonal fluctuations in the observed average age of cases was typically greater than that predicted by models, suggesting stochastic and discrete effects serve to amplify rather than obscure seasonal changes in the average age.
Since age–incidence patterns vary in a predictable manner, they may aid in the identification of unknown infections. At least one infection is thought to be involved in the aetiology of KD, but a specific agent has yet to be identified [2
]; it is possible that a previously unidentified virus is involved [6
]. Examining seasonal changes in the age and incidence of KD hospitalizations in the USA, we found that periods of high incidence corresponded to a low average age of cases, and vice versa. This pattern is in stark contrast to those exhibited by measles and rotavirus. By comparing the observed pattern to those predicted by a suite of models consistent with hypotheses about the aetiology of KD, we found that the age–incidence pattern of KD suggests the involvement of an imperfectly immunizing infection and/or an infectious agent that has a long duration of infectiousness.
For immunizing infections, both the SIR and SEIR models suggest that in order for the average age of cases to be highest during the summer/fall when the incidence of KD is low, the period of communicability would have to be long, i.e. on the order of four months or more, and mixing would have to be highly assortative. There has been considerable debate over whether KD results from an immunological cascade triggered by bacterial superantigens [25
]. Immunity following such bacterial infections is typically not life-long. For such an infection, we considered an SIRS model in which immunity wanes after 2 years. In order for cases to occur primarily in childhood, we assumed that KD is the result of an abnormal immune response that occurs upon first infection in genetically predisposed individuals and subsequent infections are not associated with symptomatic illness. We cannot discard this hypothesis for an etiologic agent with any duration of infectiousness assuming transmission-relevant mixing reflects self-reported contact patterns, which is likely the case for many respiratory infections. It is also possible that infection is imperfectly immunizing, but that cases are limited to children because of age-related susceptibilities. However, in this case, the transmission rate is inestimable without a priori
information on how risk varies with age.
Another hypothesis is that KD is caused by co-infection with two infectious agents, such as an acute viral infection that interacts with colonizing bacteria, leading to bacterial proliferation and toxin production [2
]. Data from an animal model of KD suggest that two triggers might be responsible for KD pathogenesis [28
]. Similar interactions have been hypothesized to be involved in the aetiology of invasive pneumococcal disease [29
] and meningococcal disease [30
]. We explored this using a model for co-infection with two immunizing agents, and found that the age–incidence pattern for KD is consistent with such a hypothesis provided at least one of the infections has a long period of communicability. This relationship held true for a variety of mixing assumptions, making it perhaps the most robust hypothesis. Since co-infection tends to be a rare event, the low incidence rate of KD in the USA could be consistent with co-infection among individuals with a relatively common rather than rare genetic predisposition. Knowing the prevalence of the genetic determinants of KD would help distinguish between some of the hypotheses presented here.
While there are likely age-related biases in the reporting of many diseases, including KD [32
], such biases are unlikely to affect the observed age–incidence patterns unless they vary by season. Underreporting of cases, such as failure to account for cases of ‘incomplete’ KD (in which two or more of the diagnostic criteria are not met), will create bias only if the age–incidence pattern among such cases differs from that among those in our dataset. Approximately, 15 per cent of patients may not meet the full diagnostic criteria, and these cases tend to be concentrated at the extremes of the age distribution [33
]. However, there is no evidence that seasonal patterns differ between incomplete and typical KD. Errors in coding of hospitalization records for KD may be constant throughout the year (rather than proportional to the true incidence) and independent of patient age. The influence of coding errors will be greater during periods of low incidence and may bias our estimates of the average age of cases. In this case, the bias is expected to vary by season, and therefore may confound our results. While this may contribute to the pattern observed for KD, by excluding all cases greater than or equal to 10 years of age (6.6%), the effect should be limited. Similarly, readmissions for KD may follow a non-seasonal pattern and be associated with an older average age, thereby generating a bias that could account for the observed pattern. However, an analysis of readmissions data from a subset of states in our dataset revealed that readmissions accounted for less than 10 per cent of all admissions, and 87 per cent of readmissions occurred within one month of the primary admission (see electronic supplementary material). Thus, this is unlikely to account entirely for the pattern we observed.
It would be interesting to test whether our findings are replicated in other populations. In Japan, where the incidence of KD is approximately 10–15 times higher than in the USA [34
], nationwide surveys have been conducted every 2 years since 1970 [35
]. There have been three nationwide epidemics in Japan, occurring in 1979, 1982 and 1986 [15
] and more localized outbreaks have occurred regularly since then. A shift in the age distribution of KD cases towards younger individuals during these epidemics has been noted [12
]. Furthermore, a bimodal seasonality has been observed over the past 20 years, with peaks in January and again in June/July [15
]. It would be interesting to see if the average age of cases changes in a bimodal fashion as well. If the pattern we observe in the USA also characterizes seasonal changes in the age distribution of KD cases in other countries, it could lend further insight into the aetiology.
The models we present here were parametrized and structured to represent the transmission dynamics of measles and rotavirus or to address specific hypotheses about possible etiological agent(s) of KD, and are by no means exhaustive. Other possibilities include, but are not limited to, models for strain–variable infections with complex immunity, e.g. rhinoviruses and group A streptococci. However, we believe the method proposed here based on age–incidence patterns might be applicable to other diseases with a suspected infectious aetiology, and could be used to gain a better understanding of the transmission dynamics of known infections.