Background
Rapid detection of pandemic influenza at national or regional level is a public health issue of critical importance [
1,
2]. Huge excess mortality and morbidity have been associated with the pandemics of influenza outbreaks in the past [
3]. In the aftermath of the highly pathogenic H5N1 avian influenza outbreaks worldwide [
4,
5], the growing concern [
3,
4] of a virulent form of a possible human influenza pandemic has led to the setting up of influenza surveillance systems across the globe [
6]. One of the main purposes of such worldwide expansion of influenza surveillance systems is the timely detection of influenza outbreaks of pandemic potential [
7]. The importance of timely detection lies in buying some extra time for being prepared to deal with a pandemic [
3,
8,
9]. This has also been corroborated by some recent mathematical modelling studies [
10,
11] of pandemic influenza outbreaks: a key finding suggests that there would be a small window of opportunity for getting ahead of pandemic outbreak fronts and thus helping early pandemic mitigation efforts if it could be detected early on.
Most developed countries as well as many from the developing world have some form of influenza surveillance in place [
6]. These surveillance systems are based on the reporting of disease syndromes (e.g., reports of Influenza-like illnesses (ILI)) and are generally designed to monitor levels of seasonal influenza [
12,
13]. Although the signature of pandemic influenza could be different from that of seasonal ones [
14], the traditional approach (patients presenting with clinical signs of ILI, collection of throat/nasal swab samples from some of these patients and, finally, laboratory confirmation of influenza) followed in influenza surveillance systems, in the absence of any detection algorithm applied to syndromic data, may not be able to pick it up early on. This is the reason that public health surveillance systems are being supplemented by the new state-of-the-art statistical tools [
1,
2]. The development of these new statistical tools has demonstrated the potential to automate syndromic surveillance systems, to be able to raise specific and sensitive early alerts of adverse disease outbreaks. Indeed this is a fast growing and a very active area of scientific research at the moment [
6].
At present, a number of methods [
12,
13,
15] exist to establish the onset of peak activities in the epidemics of seasonal influenza. These methods are mostly based on regression [
16,
17] or time-series [
12,
13,
15] analysis of seasonal ILI data. One such method is the Moving-Average Cumulative sums (Mov-Avg Cusum) method [
18‐
20]. Originally developed for the industrial quality control [
21], it is now frequently used for detecting the outbreaks of seasonal and pandemic influenza [
12,
22]. Recently there has been a flurry of new detection methods based on sophisticated statistical approaches [
1,
2], including those aimed at real-time monitoring and projecting of influenza cases [
23,
24]. However, challenges remain in terms of how to use the ILI surveillance data in a simple and efficient manner for timely detection of influenza pandemics.
The basic reproduction ratio R
0 (i.e., the average number of new infections produced by a single infection in a totally naïve host population) plays a central role in our understanding of infectious disease dynamics. It determines whether a new infection will successfully invade the susceptible population [
25,
26]. In the case of an ongoing epidemic, the effective reproduction ratio R replaces R
0 [
25]. In the presence of disease tracing data, R can be estimated and the in- or out-of-control status of an ongoing epidemic can be established [
27]. Where there is no availability of disease tracing data, as in influenza syndromic data, the weekly case ratio (WCR), defined as the ratio of the number of reported cases in a week to the number of cases reported in the previous week, may function as an indirect measure of R and may be suitable for raising public health alarms in the early stages of an emerging infection. Although pandemic influenza infections may grow exponentially in early invading stages (evident in the mortality data from the past pandemics [
28] or in the mathematical modelling [
10,
11,
29] of influenza pandemic), the detection algorithms so far employed in influenza surveillance systems largely ignore this natural behaviour as the basis for generating early warning of influenza outbreaks of pandemic potential.
The aim of this paper is to develop a detection algorithm, based on the estimates of WCR for expected influenza pandemics, to facilitate sensitive, specific and rapid detection of a pandemic outbreak at a regional level based on existing surveillance systems. Using the influenza surveillance data from Scotland, we first sift through the spatiotemporal patterns in the historical data by calculating WCR and N
HB, the number of health boards (HBs) that show increases in the weekly ILI cases. The joint probability distribution of WCR and N
HB is then contrasted with expected and observed patterns in the presence of pandemic influenza. As described in the next section, the expected patterns for pandemic cases are obtained from a previously published mathematical model [
29]. Observed patterns for pandemic cases are based on records from the 2008-09 season when influenza A(H1N1)v was circulating in Scotland.
We compare the performance of our detection algorithm, using simulated influenza pandemics as well as data from the 2009 influenza A(H1N1v) epidemic in Scotland, with that of the Mov-Avg Cusum method and with the ILI rate threshold method, a slightly modified form of the baseline ILI activity indicator used by the Health Protection Agency (HPA) in the monitoring of seasonal influenza in the UK.
Discussion
In this paper we compare three methods of detecting an influenza pandemic using an existing surveillance system in Scotland called SERVIS. The ILI rate threshold method uses current ILI case data to detect pandemics. This method is motivated by the current HPA's threshold levels [
30,
32] to monitor influenza activity at the national scale. The HPA thresholds serve the purpose of establishing whether seasonal influenza activity warrants some intervention (e.g., the start of antiviral prescription) of the public health departments. Mov-Avg Cusum and other variants of Cusum are already being used in public health surveillance systems [
1,
6,
12,
18]. The Mov-Avg Cusum method detects a pandemic when the cumulative number of current ILI cases is substantially higher than the expected cumulative number. The Mov-Avg Cusum statistic keeps accumulating the deviation between observed and expected values over time and when the accumulated value crosses a pre-set threshold, an alarm is triggered [
21]. It has three adjustable parameters that require optimisation for specific surveillance systems. Finally, the WCR algorithm introduced in this paper is based upon a characteristic of epidemics, their exponential growth in the early stages before control measures and depletion of susceptibles have occurred. It also assumes that pandemic influenza would occur synchronously across spatial units of influenza surveillance system in a region (as is predicted by the mathematical models [
29] for pandemic influenza in Scotland). It makes use of the joint probability distribution derived from the historical seasonal ILI data to detect a pandemic influenza. The other methods do not use any information from the data, other than to set thresholds to achieve required specificities.
The WCR algorithm appears to provide a slightly more rapid and sensitive tool for detecting of pandemic influenza - median detection time for this method ranges from 3 to 5 weeks in comparison to 4 to 6 weeks for the other two methods. Although the WCR algorithm seems to do the job more efficiently with the simulated pandemics, it performed poorly with the 2008-09 pandemic data from SERVIS. There could be several possible reasons for this poor performance. First, the 2009 influenza A(H1N1)v epidemic happened outside the normal influenza season and, second, it was mild in severity [
33,
34]. In addition, the number of SERVIS sentinel GPs in the season 2008-09 was at its sparsest level - only 20 practices, as the SERVIS system was in the process of being phased out to be replaced by a system which automatically collected data from GP systems on a daily basis. These three factors will have contributed to poor reporting of the early pandemic cases, notwithstanding the huge media coverage given to the pandemic. This is consistent with the patchiness in the reported ILI cases through SERVIS sentinel GPs between weeks 29 and 41 (Additional file
4). No method will detect a pandemic in its early weeks if the early syndromic influenza data are not reported to the surveillance system. Finally, the 2009 pandemic influenza cases were more spatially heterogeneous than those predicted by the pandemic model. (It is interesting to note that during the period April to July 2009, when there was a sentinel practice within an outbreak area, Greenock and Govanhill in the GGC HB, there was no increased reporting of ILI.) This might have contributed to the observed patchiness in the sentinel reporting.
An important aspect of our algorithm is that the detection threshold remains constant throughout. An implementation of a time-varying detection threshold could make this algorithm capable of using the seasonal ILI pattern more efficiently. In principle, this could be implemented by calculating the joint probability of (WCR, NHB) either on a week-by-week basis, or on a slightly more coarse temporal scale of the time-windows of high/low ILI activities in the seasonal data. Clearly increasing the number of sentinel GPs and the frequency of ILI case reporting would improve the temporal resolution of the WCR algorithm. In future work we will explore how many sentinel GPs are required to achieve this aim.
Furthermore, in outbreaks of a novel influenza strain, generally children and young adults of the population, who will have little or no prior immunity to the disease [
4], are disproportionately affected [
28]. Implementing our detection algorithm using these data attributes will further improve the timeliness and specificity of the detection of pandemic influenza. SERVIS data contain age attributes which could be incorporated into our algorithm, but this requires more sentinel GPs to be of use.
The WCR algorithm could be applied to any syndromic surveillance data structured by space and time. Syndromic data-sets include, but are not limited to, the triage nurse calls [
35] (e.g., the NHS24/NHSdirect calls in the UK [
36‐
38]), the over-the-counter medicine sales data [
39,
40] available in most of the developed countries, or online web search queries [
41]. These data sets are highly useful in the early detection of unusual health events [
1,
2]. Generally these data sets come with spatiotemporal attributes and, therefore, could potentially be integrated with the seasonal ILI data; this should enhance the detection process (in terms of timeliness, specificity and sensitivity) of pandemic influenza.
Conclusions
While computationally simple to implement, the WCR algorithm is capable of raising alarms, rapidly and sensitively, for influenza pandemics against a background of seasonal influenza. That it has the potential to be more specific in generating alarms for pandemic influenza could be exploited for making it more cost-effective for public health surveillance systems that collect the syndromic data at a more finer spatial and/or temporal resolution. Although the algorithm was developed using the SERVIS data, it has the capacity to be used at other geographic scales and for different disease systems where buying some early extra time is critical.
More generally, we suggest that a combination of different statistical methods should be employed in generating alarms for infectious disease outbreaks. If carefully implemented, this would provide two benefits: 1) increased sensitivity; 2) different detection methods would provide cross-checks on one another, boosting confidence in the outputs of the surveillance system as a whole.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
BKS conceived the idea of the WCR algorithm, wrote the computer codes, ran the analysis and drafted the manuscript. MEJW and NJS originally conceived the idea of developing a pandemic detection method using the existing Scottish syndromic surveillance systems, helped with the development of the WCR algorithm as well as the analysis and with drafting the manuscript. NMF provided the simulated pandemic influenza data for Scotland. CR suggested the comparison with the Mov-Avg Cusum method, helped with the analysis and provided statistical support in the development of WCR algorithm, and helped in getting access to SERVIS data. All authors have read and approved the final manuscript.