The incidence and risk factors for pdmH1N1 during the 2009 pandemic were well studied in Singapore
[18, 20–22]. In this work, cross-sectional serologic surveys suggest that about 40% of school-aged children were infected in the initial wave of pdmH1N1 in Singapore compared with 17% of adults
. These results corroborate the importance of school-aged children in influenza transmission, and the potential of schools as a source of influenza sentinel surveillance data, particularly since epidemics in school-aged children tend to lead influenza activity in older age groups
Comparison of the epidemic curves suggests that, while illness episodes from Sch-FRI reporting closely tracked those from GP ILI surveillance, laboratory confirmed pdmH1N1 cases (Sch-LCC) detected more of the earlier infections in the epidemic. This is expected since pdmH1N1 notifications were based on the prevailing national testing and reporting regime. At the start of the local epidemic from late May through June 2009, all individuals suspected of having pdmH1N1 infection were comprehensively tested as part of the Ministry of Health protocol during the containment phase of the epidemic, with this requirement abolished when the national response transitioned to mitigation phase in early July 2009 (in week 27)
. Moreover, we noticed that the distribution of pdmH1N1 notifications cases was skewed, with a few schools contributing most of the cases (data not shown) – this may be due to higher testing and reporting rates in schools with earlier outbreaks; other authors have likewise observed how cases linked to outbreaks can be over-represented in laboratory confirmed infections
For the same reasons, we found when comparing incidence rates between the systems that laboratory confirmed pdmH1N1 cases (Sch-LCC) likely detected less than 1% of all estimated infections. Since our clinic-based ILI data started only from week 25, we are unable to estimate the fraction of infections detected by laboratory confirmed cases notified earlier during the epidemic, which may have been substantially higher. However, we note that studies from other developed countries which attempt to estimate infections either by symptoms or serology likewise suggest that only a small proportion of infections are confirmed
[25–28]. Based on the comparison with Sch-FRI, temperature monitoring twice a day (Sch-DTM) may also only have identified less than one fifth of febrile respiratory illness episodes, and by extrapolation a smaller fraction of infections. Many influenza infections never result in fever
[29, 30], and those who do become febrile may not have a fever at the time of monitoring, may refrain from attending school in the first place, may take antipyretics, or may have an elevated temperature that nevertheless falls below the defined threshold; any of these circumstances would result in cases not being identified by the monitoring system. On the other hand, our novel teacher led febrile respiratory illness reporting system (Sch-FRI) covering six schools distributed across the country obtained incidence rates consistent with those observed in some school-based outbreaks where syndromic case definitions of self-reported fever and respiratory symptoms were also used
[31–33]. While data specific to pediatric populations is lacking, other studies show that symptoms occur in two thirds to three quarters, and febrile illness in about half of serologically detected infections
[18, 29, 34]. Since our ratio of illness episodes (Sch-FRI-adj) to infections was around 0.6, we suggest that self-reported FRI had detected a substantial proportion of symptomatic infections, and hence may be sufficiently sensitive as a means of detecting clusters of transmission in contrast to the other two indicators (Sch-LCC and Sch-DTM) evaluated, which may be limited in their sensitivity for triggering investigations and interventions.
A surveillance system built upon a small group of schools, as in the Sch-FRI system described here, would not allow central educational authorities to instigate responsive school closures in schools which are not enrolled in the network. However, self-reported ILI has been used successfully to investigate school-based outbreaks
[9, 33]; others have also used self-reported ILI to assess the burden of pdmH1N1 in the community and the proportions which seek care
[24, 35]. We believe that the school-based FRI reporting we describe offers some advantages over clinic-based ILI reporting: (i) it can be rapidly implemented in a centralized educational system, as in Singapore, (ii) it is not dependent on health-seeking behavior and can potentially work in areas with poor primary care coverage, (iii) it has clear denominators of the population at risk, and (iv) it does not require additional laboratory testing or serological studies. On the other hand, such systems face several challenges, including how to monitor epidemics during school holidays, the representativeness of participating schools (particularly in rural areas where transmission may be less uniform than in highly urbanized Singapore), mitigating the burden of data collection, and integrating such surveillance with data on adults and pre-school children. However, in spite of these limitations, such a system can be a useful adjunct to other more established clinic-based systems, since it is not dependent on primary care coverage or health-seeking behavior, and allows estimates of infection rates following adequate adjustment for the contribution of other causes of febrile respiratory illness and the proportion of infections that do not present with fever. Our analysis of the variation in FRI rates also suggests that FRI reporting has some potential for identifying localized transmission. We noted a wide difference in FRI rates by classrooms, with more than 10-fold difference in rates between the 5th and 95th percentile (5 vs 58 episodes per 100 children). In spite of this, FRI rates aggregated at the level of schools were relatively similar. We suggest that this apparent disconnect could be explained if we consider influenza incidence at school level as an aggregate of semi-independent self-sustaining clusters of transmission at the class-room level, which produces a wide range of cluster sizes distributed around an inherent mean. When the schools are a sufficiently large collection of classrooms (as was the case in our study, where the 6 schools had between 33 to 63 classrooms, with a median class size of 31 students and inter-quartile range from 29 to 39), then the school level FRI rate reflects the average size of a transmission cluster in the classroom setting. Cauchemez et al. have demonstrated that, within the school environment, classroom level transmission dominates
, and our study adds to the emerging evidence that this is indeed the case. Notably, in the Singapore school system, students mostly interact within the same class, with most classes conducted within the same room throughout the day for lessons, and this may have accentuated the effect. Additional studies will be needed to clarify the pattern of influenza transmission within schools, as this will have substantial implications on control measures, since execution of closures and interventions at classroom level, if effective, would be far less disruptive than equivalent measures at the level of entire schools or even all schools within geographic areas. However, if there is intent to intervene using such data, then febrile respiratory illness may have to be monitored in real-time and followed-up by confirmatory testing of students identified, which may be logistically challenging since some students would be absent from school at the time of their illness; this may also be expensive if implemented at a national level. There may also be issues with variations in data quality if deployed on a wider scale or for longer periods, and as such FRI reporting may function best either when used for short periods such as for detecting transmission clusters during severe epidemics, or in sentinel schools with dedicated support staff to ensure that it is properly collected when used for long-term surveillance of influenza activity. Finally, modeling studies should be attempted to suggest appropriate triggers for interventions (such as a certain number of FRI episodes within a particular time frame), and the potential effect of any interventions on reducing influenza transmission.
Limitations of our study include the fact that the different indicators were collected over different time periods, especially for daily temperature monitoring which was only available for a limited period according to the Ministry of Education mandate. Furthermore, it is not possible on the basis of our analysis alone to determine which of the different systems most accurately reflected the true timing of the epidemic, especially in view of data availability, reporting lags and day of week effects which limited our resolution to weekly incidence and may have introduced bias from one or more datasets. Moreover, missing onset dates from about a quarter of the reported FRI episodes may have biased the epidemic curve for this dataset. Ideally, we would also have wanted information on all presentations of acute respiratory illness, with additional information on the recorded temperature, and not just FRI; however, this would have increased the complexity of data collection, and we therefore adopted the compromise of using just febrile respiratory illness as our case definition for identifying possible pdmH1N1 infections. In addition, we had inadequate data on the baseline incidence of FRI, and this may have biased our estimation on the contribution of non-pdmH1N1 causes to FRI. Notably, one study from the US suggests a substantially higher baseline incidence of self-reported febrile respiratory illness in children (ages < 18 yrs), although the case definitions used in that study were slightly less specific
. Some routine collection of baseline FRI incidence may hence be necessary to aid interpretation of data from FRI monitoring systems during epidemics and pandemics. Also, the serological survey that was used for comparison had a post-epidemic sample which was taken from 1 October 2009 to 2 June 2010. Subsequent smaller waves of H1N1 2009 within that period
 and waning of initial antibody levels
[38, 39] might have affected the results. It must also be noted that the data used to adjust for the sensitivity of the HI assay was based on results from adults
, since we did not have data specific to younger populations. However, we note that the estimates of infection rates in our study are fairly similar to other serologic surveys of corresponding age groups in other countries following a single epidemic wave
[1, 26, 28, 40–42]. Finally, our study lacked the necessary data such as differences in interventions or time lag between symptom onset and sickness absenteeism to explain our intriguing finding on the large variation in clinical attack rates at classroom level.