Click on the links below to access all the ArticlePlus for this article.
Please note that ArticlePlus files may launch a viewer application outside of your web browser.
Annual influenza epidemics occur worldwide, resulting in considerable morbidity, mortality, and economic burden.1 The incidence of influenza displays a seasonal pattern in temperate areas, with marked peaks in winter (December–April in the northern hemisphere and June–September in the southern hemisphere).1–3 Despite this regular occurrence, the onset, duration, and magnitude can differ widely. Little is known regarding the mechanism of geographic spread, and even the reasons for the seasonality of influenza are unclear. Better understanding of those differences might help to predict or prevent these epidemics.2,4
Several descriptive studies have explored the spatiotemporal patterns of influenza. Viboud et al5 analyzed epidemic synchrony across Australia, France, and the United States between 1972 and 1997 and found a high correlation between the effect and the timing in the latter 2 countries. Bonabeau et al6 modeled influenza spread within France (1988–1995) and showed that the illness diffused rapidly across the country, with the occurrence of homogeneous global mixing before local epidemics peaked. Mugglin et al7 modeled patterns in Scotland during 1989–1990 and identified 2 clusters in the northern and southern regions of a high-population corridor. Paget et al8 found evidence for a west-to-east spread of peak influenza activity across 20 European countries in 3 of 5 recent winters (1999–2004). Sakai et al9 used kriging and simple statistics to characterize influenza-like illness patterns across Japan from 1992 to 1999 and found nationwide epidemic patterns that spread in concentric circles from western-central to eastern regions. Although these findings help to explain the broad spatiotemporal patterns of influenza, they may not be adequate for more local outbreaks.
Influenza outbreaks impose huge infrastructure demands on healthcare systems and exact economic costs through sickness-related absenteeism, disrupted work schedules, and lost productivity. Strategies to prevent or contain influenza outbreaks can be categorized into 2 broad groups: antiviral vaccines and nonpharmaceutical measures (eg, school or workplace absenteeism, restrictions on public activities). In relation to nonpharmaceutical measures, public health officials are required to evaluate the severity of local disease clusters and to develop mitigation strategies. To this end, it is essential to evaluate changes in the spatial and temporal patterns of influenza epidemics based upon detailed information such as the location and time of incidence.
We used space-time permutation scan statistics to identify the spatiotemporal patterns of influenza epidemics in Fukuoka, Japan, during the 2006–2007 transmission season. We identified and characterized a wave of influenza epidemics across Fukuoka and discuss the potential factors associated with influenza's intriguing pattern of spread.
In Japan, the systematic surveillance of influenza as a notifiable disease began in 1981 under the Infectious Disease Control Law. This system, organized by the Ministry of Health and Welfare, involves about 5000 sentinel medical institutions (mainly pediatric hospitals and clinics), accounting for about 8% of the total number of pediatric hospitals and clinics throughout the entire country.10 A case of influenza is defined by a sudden fever (>38°C), respiratory symptoms, general malaise, and myalgia. The number of influenza patients is reported on a weekly basis from 198 sentinel medical institutions within Fukuoka Prefecture, southwest of Tokyo, Japan (eFig. 1, available with the online version of this article). The number of sentinels is based on the population of the area in which the health center is located: A health center with a population of <75,000 would have 1 sentinel, an area with a population of 75,000–125,000 would have 2, and an area with a population of >125,000 would have 3 or more according to the formula (population − 125,000/100,000).11 Sentinels are recruited on a volunteer basis and are asked to send clinical data to the Fukuoka Prefecture Medical Association, the Fukuoka Prefectural Government, and the Fukuoka Institute of Health and Environmental Sciences (the municipal public health institute of Fukuoka Prefecture). Data are reported electronically to the Infectious Disease Surveillance Center at the National Institute of Infectious Diseases in Tokyo, Japan.12
We analyzed data only for Fukuoka Prefecture. These data were obtained from the National Epidemiological Surveillance of Infectious Diseases system, which monitors infectious disease events among the approximately 5 million residents of Fukuoka Prefecture. Addresses were matched to sentinel medical institutions for 113,503 influenza cases from 29 August 2005 to 3 June 2007. All cases were geocoded according to latitude and longitude. The Fukuoka Prefecture Environmental Health Research Advancement Committee approved this study on 17 December 2004 (reference number, 16–1803).
SaTScan software analyzes spatial, temporal, and space-time data by using spatial, temporal, or space-time scan statistics.13,14 SaTScan is a trademark of Martin Kulldorff and was developed under the joint auspices of Martin Kulldorff, the National Cancer Institute, and Farzad Mostashari at the New York City Department of Health and Mental Hygiene. The space-time scan statistic is used to determine which clusters merit further investigation and which are the result of chance alone.15 SaTScan identifies a cluster at any location of any size up to a set maximum, which limits the problem of multiple statistical tests.
The space-time data could be analyzed in either retrospective or prospective fashion. In a retrospective analysis, the analysis is done only once for a fixed geographic region and a fixed study period. SaTScan software scans over multiple start dates and end dates, evaluating both live clusters (lasting until the study period and date) and historic clusters (that ceased to exist before the study period end date). The prospective analysis is used for the early detection of disease outbreaks, when analyses are repeated every day, week, month, or year. Only live clusters—clusters that reach all the way to current time as defined by the study period end date—are then searched for.
We used the prospective space-time permutation scan statistics13 for the early detection of influenza outbreaks. Because outbreaks are indicated by the number of cases, there is no need for population-at-risk data. This statistic makes minimal assumptions about time, geographic location, and the size of an outbreak, and it adjusts for natural, spatial, temporal, and space-time variation. The space-time permutation scan statistic uses a 3-dimensional cylindrical window while scanning. The base of the cylinder represents space, and the height represents time. Both the circular geographic base and the start date, which are mutually independent, are flexible. The likelihood ratio test statistic is constructed using a computational algorithm to calculate the likelihood of each window in 3 dimensions.
SaTScan has been applied to various diseases in a number of areas. We applied SaTScan to data collected in Fukuoka Prefecture, using a circle as the base of the scanning cylinder. SaTScan was used to identify clusters of sentinel medical institutions in Fukuoka Prefecture, and the scanning parameters were set to search only for areas with high proportions of influenza. Because geographic overlap was not used as a default setting, secondary clusters do not overlap the most significant cluster. For space-time permutation model, the maximum cluster size is defined as a proportion of number of cases. To scan from small to large clusters, the maximum cluster size was set at 50% of the total number of cases in the scanning window.
The choice of maximum temporal window size was determined by the nature of the surveillance setting. For prospective surveillance the goal is to detect ongoing outbreaks. Early detection of disease outbreaks enables public health officials to implement disease control and prevention measures at the earliest possible time.10,16,17 For an infectious disease, improvement in detection timeliness by even 1 day might aid public health officials in controlling the disease before it becomes widespread. However, making the temporal window too small can substantially reduce the power to detect slowly emerging disease outbreaks. At the same time, these methods are meant for the rapid detection of disease outbreaks, and, depending on the disease, late detection of an outbreak might not provide public health benefit. For these reasons, the maximum temporal cluster size was set at 14 days for the early detection of outbreaks using weekly data.
Monte Carlo simulations, generating random replications of the dataset under the appropriate null hypothesis, are used to determine the statistical significance of these results. The P values for these tests are calculated by comparing the rank of the maximum likelihood from the real dataset with the maximum likelihoods from the random datasets with P = rank/(1 + number of simulations).18 The number of replications should be a minimum of 999 to ensure adequate precision. However, 9999 replications are recommended when computing time is not an issue.18 Thus, we used 9999 Monte Carlo replications to estimate the significance levels of these clusters.
For all analyses, we report the most likely and secondary clusters with statistical significance of P < 0.02, with a recurrence interval of 1 year (52 weeks). Based on the size of the log-likelihood ratio, we identified the most likely and secondary clusters. The window with the maximum likelihood is the most likely cluster, that is, the cluster least likely to be due to chance. There will almost always be a secondary cluster that is almost identical with the most likely cluster and that have almost as high likelihood value, because expanding or reducing the cluster size only marginally will not change the likelihood very much. A large number of less likely overlapping clusters can be found with high significance around any cluster as the inclusion/exclusion of a small population may not have a large impact on the results.18 Thus, when mapping the detected clusters, we report only the most likely nonoverlapping clusters to simplify the presentation of results.
Because we are searching for live clusters by using prospective analysis, only the clusters in the most recent season are detected. To this end, we performed the prospective space-time permutation scan statistics during the 2006–2007 transmission season.
Between 29 August 2005 and 3 June 2007, there were 2 seasonal peaks in the weekly time series of the total number of cases reported by the sentinel clinics/hospitals in Fukuoka Prefecture (Fig. 1). Influenza displays a seasonal pattern in temperate areas, with marked peaks in winter.1–3 In each year at all reporting sites, the annual influenza season began in November or December, peaked during January or February, and returned to baseline between April and June. Seasonal peaks in influenza infection occurred once a year. Nationwide epidemics lasted for 3 to 4 months, but successive or overlapping waves of infections with influenza types A and B sometimes resulted in a more prolonged outbreak in the 2005–2006 transmission season.
We tested the space-time permutation scan statistics using the weekly prospective analysis of influenza surveillance data from 29 August 2005 to 3 June 2007 to identify statistically significant space-time influenza clusters (eTable 1, available with the online version of this article and Fig. 2). During the 2006–2007 transmission season, the earliest “most likely” cluster was first detected in Fukuoka City from 20 to 26 November 2006 (eTable 1, Fig. 2). This cluster showed 3 cases, when only 0.13 could be expected theoretically (relative risk [RR] = 22; P = 0.019; radius = 1.67 km).
The next highest risk cluster during the 2006–2007 transmission season was detected in the Chikushi area, a suburb of Fukuoka City, from 4 to 10 December 2006 (eTable 1, Fig. 2C). This cluster showed 13 cases when only 0.40 were expected (RR = 33; P = 0.0001; radius = 3.69 km). It is unlikely that such cluster of cases was caused by random variation. In the early stages of the influenza season (between 20 November 2006 and 14 January 2007), several statistically significant clusters with comparatively high RR were detected in the urban regions and neighboring areas (eTable 1, Fig. 2A). Several small-scale clusters (radius <10 km) with comparatively high RR were detected for several consecutive weeks (eTable 1, Fig. 2A). Middle-scale clusters (radius =10–30 km) also appeared gradually as the season progressed (eTable 1, Fig. 2B). Toward the end of the season, we observed large-scale clusters (radius ≥30 km) continuously, lasting for several weeks (eTable 1, Fig. 2C). The largest cluster appeared from 5 March 2007 to 11 March 2007 and encompassed 121 sentinel clinics/hospitals (eTable 1, Fig. 2C). This outbreak comprised 10,309 cases when only 8988 were expected (RR = 1.15; P = 0.0001; radius = 50.09 km). These middle- to large-scale epidemics persisted for about 6 weeks and then gradually waned (eTable 1, Fig. 2D).
We used space-time permutation scan statistics to examine the spread of influenza in Fukuoka Prefecture, based on minimal assumptions about the spatiotemporal characteristics of influenza clusters. Our results suggest that geographic factors affect disease spread. The influenza virus first appeared in Fukuoka City early in the season (from November to December 2006). Small-scale (radius = 1.75–5.04 km) and high relative risk clusters (RR = 14–33) were detected in the Fukuoka City and Chikushi areas for several weeks (eTable 1, Fig. 2A). Fukuoka City is the most densely populated urban area in Fukuoka Prefecture and is a hub for road, rail, and air travel and transport.19 Thus, our results suggest the importance of human movement and public transportation in the spread of influenza, consistent with previous reports.5,6,9,20
Several weeks after the first clusters were detected in Fukuoka City, significant clusters (radius = 0–17.38 km) were observed in Kitakyushu City (a secondary urban area in Fukuoka Prefecture) and the Onga and Munakata areas (suburbs of Kitakyushu City) (eTable 1, Fig. 2B). Kitakyushu City is also the gateway to one of the major Japanese islands, Honshu.19 These results suggest that early influenza outbreaks spread between urban centers and reiterate the importance of both population and traffic density in spreading influenza.5,6,9,20
Small- to middle-scale clusters (radius = 0–17.38 km) persisted in the Fukuoka and Kitakyushu urban areas for several weeks, after which the clusters gradually diffused and middle- to large-scale clusters (radius = 15.95–50.09 km) developed in rural areas (eTable 1, Fig. 2C). These results show that influenza outbreaks increased within urban areas before spreading along main roadways and railways to initiate large outbreaks in rural areas throughout Fukuoka Prefecture.
Finally, once a statistically significant outbreak cluster was identified, the peak occurred within 16 weeks (112 days). Influenza outbreaks took place over about 28 weeks (196 days) from 20 November 2006 to 3 June 2007. These results are important for contingency planning. In a pandemic situation, the illness would be expected to spread more rapidly.21–24 Further investigation into these geographic trends in the spread of epidemic foci may have implications for the control of outbreaks.
Understanding local spread is an important step in refining the prediction of the spatiotemporal dynamics of influenza epidemics.25,26 Our results suggest that influenza outbreaks start in cities, persist for a short period in the urban area, and then disperse throughout rural areas. These results demonstrate the potential of scan statistic to identify epidemiologically valid outbreaks that may be overlooked by standard surveillance. These types of clusters may help to identify risk factors associated with outbreaks, to develop mitigating strategies for an influenza outbreak, and to reveal differences among health units in terms of reporting or detecting outbreaks.
Our study has several methodologic limitations. First, cluster studies are subject to their own set of limitations. Many of these studies are based on surveillance data; therefore, reporting bias may complicate results. This bias can occur anywhere in the reporting chain, from the initial tendency of a patient to seek health care to the recording of the case in the disease registry. Because underreporting may vary among jurisdictions, a cluster study may actually reflect differences in reporting rather than incidence among geographic regions.27,28 In addition, there is controversy regarding the best method of presenting the results of scan statistics.29 Here, we presented the most likely statistically significant clusters. Finally, there is always an issue with the reliability of geocoding. Geocoding is affected by the resolution of specific information in the database and variation in the incubation period of the disease, which may result in uncertainty regarding the location of exposure to an infectious agent.
Second, our results might include potential biases because sentinel medical institutions were recruited on a voluntary basis. For applying the space-time permutation model to detect disease outbreak using the surveillance monitoring data, further discussion of disease-specific issues would be important.
Third, the geographic boundary of the detected cluster is not necessarily the boundary of the true cluster. Although we used a circle as a base for the scanning cylinder, other scanning window shapes also occur,30 and circular scan statistics can detect noncircular cluster areas.31 Although all detected clusters were approximately circular, infectious diseases can assume other cluster types, and more complicated clusters such as oval clusters may be more realistic.32
In summary, disease dissemination is a complex function of host movement and the age structure of susceptibility. Our findings are consistent with previous studies with regard to the effects of population mobility; however, we were not able to examine the effects of individual susceptibility. An investigation of the role of social and demographic factors in the spread of influenza would require not only disease data but also data on individual movements in Japan. Because influenza is likely spread through both human movement and social transmission networks, these will be critical topics for future study.
We thank the Fukuoka Prefectural Government, Disease Prevention and Health Promotion Division.
1. Simonsen L. The global impact of influenza on morbidity and mortality. Vaccine
. 1999;1(suppl 17):S3–S10.
2. Cox NJ, Subbarao K. Global epidemiology of influenza: past and present. Annu Rev Med
3. Viboud C, Alonso WJ, Simonsen L. Influenza in tropical regions. PLoS Med
4. Cox NJ, Subbarao K. Influenza. Lancet
5. Viboud C, Boelle PY, Pakdaman K, et al. Influenza epidemics in the United States, France, and Australia, 1972–1997. Emerg Infect Dis
6. Bonabeau E, Toubiana L, Flahault A. The geographical spread of influenza. Proc Biol Sci
7. Mugglin AS, Cressie N, Gemmell I. Hierarchical statistical modelling of influenza epidemic dynamics in space and time. Stat Med
. 2002;21: 2703–2721.
8. Paget J, Meerhoff T, Meijer A. Five winters of influenza activity in Europe: an evaluation of the indicators used to measure activity and an assessment of the timing, length and spread of influenza. The Ninth European Programme for Intervention Epidemiology Training (EPIET) Scientific Seminar; 2004.
9. Sakai T, Suzuki H, Sasaki A, et al. Geographic and temporal trends in influenza like illness, Japan, 1992–1999. Emerg Infect Dis
. 2004;10: 1822–1826.
10. Hashimoto S, Murakami Y, Taniguchi K, et al. Detection of epidemics in their early stage through infectious disease surveillance. Int J Epidemiol
11. Murakami Y, Hashimoto S, Taniguchi K, et al. Distributions of the numbers of monitoring stations in the surveillance of infectious diseases in Japan. Nippon Koshu Eisei Zasshi
12. Okabe N, Yamashita K, Taniguchi K, et al. Influenza surveillance system of Japan and acute encephalitis and encephalopathy in the influenza season. Pediatr Int
13. Kulldorff M, Heffernan R, Hartman J, et al. A space-time permutation scan statistic for disease outbreak detection. PLoS Med
14. Kulldorff M, Nagarwalla N. Spatial disease clusters: detection and inference. Stat Med
15. Kulldorff M, Athas WF, Feurer EJ, et al. Evaluating cluster alarms: a space-time scan statistic and brain cancer in Los Alamos, New Mexico. Am J Public Health
16. Goldenberg A, Shmueli G, Caruana RA, et al. Early statistical detection of anthrax outbreaks by tracking over-the-counter medication sales. Proc Natl Acad Sci U S A
17. Lazarus R, Kleinman KP, Dashevsky I, et al. Using automated medical records for rapid identification of illness syndromes (syndromic surveillance): the example of lower respiratory infection. BMC Public Health
18. Kulldorff M and Information Management Services Inc SaTScan
v 7.0 [computer program]. Software for the spatial and space-time scan statistics. Available at: www.satscan.org
19. The North Kyushu Region Transportation Planning Commission. North Kyushu Region Person Trip Survey The North Kyushu Region Transportation Planning. The North Kyushu Region Transportation Planning Commission; 2006.
20. Viboud C, Bjornstad ON, Smith DL, et al. Synchrony, waves, and spatial hierarchies in the spread of influenza. Science
21. Germann TC, Kadau K, Longini IM Jr, et al. Mitigation strategies for pandemic influenza in the United States. Proc Natl Acad Sci USA
22. Ferguson NM, Cummings DA, Cauchemez S, et al. Strategies for containing an emerging influenza pandemic in Southeast Asia. Nature
. 2005;437:209 –214.
23. Ferguson NM, Cummings DA, Fraser C, et al. Strategies for mitigating an influenza pandemic. Nature
24. de Jong JC, Claas EC, Osterhaus AD, et al. A pandemic warning? Nature
25. Ferguson NM, Galvani AP, Bush RM. Ecological and immunological determinants of influenza evolution. Nature
26. Grenfell BT, Pybus OG, Gog JR, et al. Unifying the epidemiological and evolutionary dynamics of pathogens. Science
27. Voetsch AC, Angulo FJ, Rabatsky-Ehr T, et al. Laboratory practices for stool-specimen culture for bacterial pathogens, including Escherichia coli
O157:H7, in the FoodNet sites, 1995–2000. Clin Infect Dis
. 2004; 3(suppl 38):S190 –S197.
28. Flint JA, Dore K, Majowicz SE, et al. From stool to statistics: reporting of acute gastrointestinal illnesses in Canada. Can J Public Health
. 2004;95:309 –313.
29. Boscoe FP, McLaughlin C, Schymura MJ, et al. Visualization of the spatial scan statistic using nested circles. Health Place
30. Kulldorff M, Huang L, Pickle L, et al. An elliptic spatial scan statistic. Stat Med
. 2007;25:3929 –3943.
31. Kulldorff M, Zhang Z, Hartman J, et al. Benchmark data and power calculations for evaluating disease outbreak detection methods. MMWR Morb Mortal Wkly Rep
. 2004;53(suppl):144 –151.
32. Onozuka D, Hagihara A. Geographic prediction of tuberculosis clusters in Fukuoka, Japan, using the space-time scan statistic. BMC Infect Dis