Spatiotemporal analyses of telehealth calls may serve as a useful adjunct to influenza surveillance. Our analyses of NHS Direct fever calls have provided a clear and detailed geographical description of seasonal influenza in England in 2005–2006. We have identified two distinct periods of significantly high NHS Direct fever calls regarding school-age children. The first rise originated in North-West England during late November 2005 and spread in a predominantly eastwards and then southwards direction over the following month. The second rise began in Central England during mid-January 2006 and moved southwards to eventually cover all of southern England. The timing and geographical location of these rises in fever calls appear similar to a national influenza B outbreak occurring during winter 2005–2006 affecting the same age group, although they were visible slightly earlier than in other data sources. We also identified significantly elevated numbers of vomiting calls in Central and South-East England during December 2005 to February 2006, within a longer period of elevated norovirus outbreak reports in the South-East. There was no apparent spatial correlation between vomiting calls and an April 2006 rise in norovirus outbreaks in the North-East of England. Assessment of the usefulness of vomiting call data for monitoring the spread of norovirus was limited by inconsistent regional level outbreak sample data within our norovirus data source.
We discuss these results from two perspectives. The first and general perspective relates to the potential of a telephone triage system, such as NHS Direct, for surveillance purposes; the second perspective for discussion concerns the specific findings and limitations of the current spatial analysis.
The potential of telephone triage systems for surveillance
Telephone triage health systems may be useful for surveillance purposes for a number of reasons. These services are generally available to the entire population of the area covered and, in the case of NHS Direct, are centrally managed. This means that clinical algorithms are chosen in a consistent manner between call centres (although experienced NHS Direct nurses may be more likely to provide self-care advice than more junior colleagues [
26]). Second, the data are available on a daily basis as opposed to GP- and laboratory-based surveillance systems, which generally report weekly. Third, telehealth is many people's first or only point of contact with the health service, providing an early opportunity to identify an increase in illness potentially 'under the radar' of other systems.
Despite the above, call data will only provide meaningful intelligence if callers' reported illness is representative of community morbidity. In England and Wales, total NHS Direct call rates (82 per 1000 in England) are approximately 3% of annual GP consultation rates for all illness (2700 consultations per 1000; A. Elliot, personal communication). One-quarter of the population have used NHS Direct [
27]. Our work identified areas of significantly high NHS Direct call rates in Northern England and substantial variation between PCTs (ranging from 33 to 212 calls per 1000 people per year). This disparity is influenced by a combination of factors, including local arrangements for provision of GP out-of-hours services (generating on average 160 calls per 1000 people per annum [
28]), social deprivation [
29] and the length of time the local NHS Direct site has been in operation [
27].
NHS Direct receives a disproportionately high number of calls about children under 4 years old (four times higher than the total call rate) and women of child-bearing age (twice the male call rate). Both of these groups traditionally have a high health service utilisation in the UK [
30]. This reporting tendency could have advantages for timely sentinel surveillance of children, to provide early warning of more widespread population impacts. Older people, who have high GP and hospitalisation rates, use NHS Direct the least of any age group. This age and gender bias is consistent with large public telephone triage systems in New Zealand [
31], Australia [
32] and Canada [
33]. NHS Direct data are, therefore, most suited to surveillance of common illnesses in the UK for those aged below 65 years. Measures must be taken to account for spatial variation in usage, such as establishing local baselines from which to detect real variation in call incidence (as employed in this study using a scan statistic).
Prospectively NHS Direct telephone triage data have been used to identify sharp rises in syndromes at a regional and national level [
16,
17]. The retrospective analysis reported here has demonstrated that it is also possible to identify a sub-regional rise in calls suggestive of influenza infection. Other potential benefits include the use of the data to inform the NHS Direct responsive messaging service, an operational tool designed to help manage service demand and to inform patients of relevant topical information (for example, 'flu and colds', 'diarrhoea and vomiting'). Trends in calls may either act as a trigger to initiate responsive messages or provide statistical support for the ongoing provision of this health advice. In addition, as well as monitoring calls suggestive of infections, these data are also used for the acute response to major incidents. For example, during July 2007, daily surveillance of calls identified a statistically significant 40% rise in calls in the Gloucester flood region (caused in part by a rise in people seeking health information and advice). This information was used to brief a national incident team and the Government on the health effects of the floods.
A qualitative evaluation has explored the usefulness of these telephone triage data for surveillance amongst a sample (
n = 91) of users of the system (those that receive surveillance bulletins and 'alerts') [
34]. The most commonly cited benefits of the surveillance were: providing real-time information for incident teams, local PCTs, media messages, and question and answer briefings for GPs and the public; supporting national plans (for example, surveillance of heat stroke calls during the heatwaves of 2003 and 2006); and preparing laboratory staff for an expected increase in specimens.
In recent years, the number of telephone triage systems has increased internationally in reaction to demands for rapid assessment and reassurance of people with health problems, and in an attempt to reduce the use of face-to-face health services, especially out of hours. Therefore, our results may be useful in other countries where there is a perceived need for real-time monitoring of common syndromes (both infectious and non-infectious). The ability of the telephone triage reporting software to map call data in real-time (not the case with NHS Direct) would also prove useful for analysing local demand for telephone triage, and for evaluating the impact of promotional activities.
Strengths and weaknesses of the current spatial analysis
There are several weaknesses to this work. NHS Direct vomiting calls are likely to be caused by a range of viral and bacterial disease pathogens. This may partially hide, within the data, the true seasonality of norovirus reported to NHS Direct. It is possible that only new norovirus variants will have a substantial impact on the national burden of vomiting calls for this age group (≥ 5 years) displaying a pattern of national diffusion rather than regional raised incidence. For example, the novel norovirus genogroup II4 variant caused an increase and atypical summer peak in outbreaks across the UK and continental Europe [
15] during 2002. Also, norovirus-infected individuals phoning NHS Direct may be reflective of norovirus outbreaks in community settings (for example, schools, food outlets), which exhibit little seasonal variation [
6] and therefore do not exhibit clear seasonality within our telehealth data. Conversely, outbreaks in semi-closed institutional settings (for example, hospitals, residential homes) show a clear winter peak but these cases are unlikely to telephone NHS Direct because they are already receiving care. This work has not studied the spatiotemporal variation in vomiting calls about young children (in which there is the highest incidence of norovirus infection [
5]). Calls about this group were excluded to remove the potentially confounding affect of rotavirus, which peaks during March in the UK. A study of the epidemic behaviour of rotavirus in the US demonstrated an annual South-West to North-East movement across the country [
35]. Analysis of NHS Direct vomiting and diarrhoea calls about young children is required to explore further the spread of viral gastroenteritis in the UK, and respond to calls to supplement existing surveillance systems for infectious intestinal disease [
14]. Unfortunately the norovirus outbreak data in our study does not represent outbreaks nationally and data from alternative laboratory or outbreak reporting systems will be sought for future work.
Common respiratory viruses, such as respiratory syncytial virus (RSV), may have contributed to the rise in fever calls (although RSV predominantly effects children younger than 5 years). Previous work estimating the contribution of a range of respiratory pathogens to NHS Direct calls found that RSV was responsible for approximately 15% of NHS Direct 'cough' calls [
36]. Although a significant relationship was found between the seasonal variation in fever calls and influenza, no significant relationship was observed with RSV, parainfluenza or rhinovirus.
These analyses used only 151 relatively large spatial units (PCTs) to identify the approximate areas of rises in fever calls, rather than smaller units (postcode areas) to identify local disease transmission or hierarchical diffusion of disease from large to smaller urban centres. The lack of spatial precision means that these results are subject to the modifiable areal unit problem (MAUP) [
37] whereby arbitrary units are used for spatial reporting. Our analysis suffers from what Armhein called the 'scale effect' of the MAUP [
38]; the grouping of small areas into larger ones (postcode districts into PCTs). This had the effect of reducing the number of spatial units, increasing the average number of cases within each unit, but reducing the variation in incidence between units. We may have detected significant clustering at a sub-PCT level by using the smaller postcode district, although there were only 0.7 vomiting and 0.3 fever calls per week per postcode district during our study period. It is yet to be demonstrated that syndromic surveillance systems can
prospectively and
consistently provide early warning of localised outbreaks of disease. Their utility continues to be seen on a city-wide or regional basis at best. This study has used syndromic (proxy) measures of influenza and norovirus to identify a syndromic pattern consistent with contagious influenza diffusion at a PCT level, rather than onward transmission from an index case or single outbreak. PCTs are a relevant spatial unit for syndromic surveillance as primary care is likely to be among the first services to experience the effects of epidemic outbreaks.
The scan statistic employed here has previously been used for analyses of specific infections (for example, shigella [
22] and
E. coli [
24]), ongoing public health surveillance (for example, New York City [
2]), and monitoring disease during major sporting events (for example, the Kentucky Derby Festival [
39]). Our work is novel in the use of the scan statistic for surveillance of telehealth calls, a data source that is infrequently used for syndromic surveillance purposes [
40]. Although many NHS Direct calls reflect self-limiting disease (for example, influenza B, norovirus) this illness is still disruptive on a societal level and may go unreported by other healthcare surveillance systems. This work, for the first time, provides evidence that it is possible to describe the diffusion of national influenza outbreaks using telehealth data. When using the scan statistic, the underlying spatial variation in incidence is unimportant until it reaches a significant threshold level, clusters are detected, and an alert may be issued to public health teams. Other statistical approaches to deriving these critical values have been described [
41,
42]. An alternative approach, however, is kriging, where the entire range of disease incidence across a geographical area is modelled. Kriging uses interpolation techniques to estimate unknown point values of disease incidence from surrounding known point values, from which disease contours are generated [
43]. The method has been employed successfully to describe the spread of the peak in influenza across Japan [
44] and Europe [
45]. For prospective surveillance, however, peak weeks are not known in advance and any global spatial clustering caused by differences in case definitions or reporting behaviour between regions would produce misleading results. The space-time scan statistic accounts for these spatial differences by using a control period from which to establish local baselines for each spatial unit. It may therefore be considered for future national surveillance initiatives, for example, the proposed hospital-based surveillance and 2012 Olympic planning in the UK.
Spatiotemporal analysis of NHS Direct syndromic data should supplement rather than replace the current analyses of regional call data [
16] and other components of the UK influenza surveillance programme. It may help to identify sub-regional variation in influenza rates that may not manifest themselves within data from sentinel networks of GPs covering only a proportion of the population [
46]. The added value demonstrated by our retrospective analyses was the early identification of a local rise in fever calls; a syndromic 'signal' that may require further investigation via conventional public health means. Used prospectively our analyses could be used to: warn hospitals, doctors, schools and nursing homes of impending problems; identify areas to introduce or enhance microbiological sampling; and describe geographic variation in influenza incidence throughout an epidemic period. We used fever calls regarding school-age children because we were studying a national influenza B outbreak. Spatial analyses of data about the NHS Direct 'colds and flu' syndrome are already used for national surveillance of ILI in adults [[
16,
47]]. Further winters' spatial analysis of respiratory syndromes is recommended to assess operational reliability of these results, and also to monitor influenza A strains associated with a high adult incidence.