Detection of epidemics in their early stage through infectious disease surveillance

Shuji Hashimotoa, Yoshitaka Murakamib, Kiyosu Taniguchic and Masaki Nagaid

a Department of Epidemiology and Preventive Health Sciences, School of Health Sciences and Nursing, University of Tokyo, Hongo 7–3–1, Bunkyo-ku, Tokyo 113-0033, Japan. E-mail hasimoto{at}epistat.m.u-tokyo.ac.jp
b Division of Health Informatics and Biostatistics, Oita University of Nursing and Health Sciences, Japan.
c Division of Intelligence and Policies, Infectious Disease Surveillance Centre, National Institute of Infectious Diseases, Japan.
d Department of Public Health, Saitama Medical School, Japan.


    Abstract
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Background Surveillance of infectious diseases is done in many countries. The aims of such surveillance include the detection of epidemics. In the present study, the possibility of detecting an epidemic in its early stage using a simple method was evaluated for 16 infectious diseases.

Methods We used as an index the number of cases per week per sentinel medical institution in the area covered by a health centre in infectious disease surveillance in Japan in 1993–1997. Periods of epidemics in health centre areas were determined according to the reported indices. The simple method used for detecting the early stage of an epidemic is that if the index exceeds a critical value, then an epidemic will begin in the following 4 weeks. The sensitivity, specificity and positive predictive value for this epidemic warning were evaluated for given critical values.

Results When the specificity of the epidemic warning was more than 95%, the sensitivity was more than 60% in ten diseases, and more than 80% in four diseases (influenza-like illness, rubella, hand-foot-and-mouth disease, and herpangina). The positive predictive value was between 15.6% and 31.4% in these ten diseases.

Conclusion The early stage of epidemics of some infectious diseases might be detectable using this simple method.

Keywords Infectious disease, surveillance, epidemic

Accepted 13 April 2000


    Introduction
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Surveillance of infectious disease is conducted in many countries.1–9 The aim of such surveillance includes detection of epidemics, especially detection in the early stage, essential for the control of epidemics of infectious diseases.10–11 Epidemics generally begin in small areas and subsequently spread widely. In order to detect an epidemic in its early stage, observation of the data for small areas is important.

Many epidemiological and statistical studies have investigated the detection of epidemics through surveillance.2,5,11–15 These investigations have mainly focused on epidemics of influenzalike illnesses in a large area (whole country), while epidemics in small areas (such as city, town) and of other infectious diseases have received insufficient attention.11,14,15 Several methods for detecting and forecasting epidemics, such as time series analysis, were applied. Some methods require strict assumptions in their underlying models and are too complicated to be applied to the data for small areas.13,15 We think that simple methods for detecting epidemics would be more practical for public health activities.

In this study, the possibility of detecting epidemics in their early stage in small areas was evaluated for 16 infectious diseases using a simple method based on data from infectious disease surveillance in Japan in 1993–1997.


    Materials and Methods
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Surveillance of infectious disease in Japan
The surveillance of infectious diseases in Japan has been described elsewhere.9 The paediatric disease surveillance system is explained below. This system, organized by the Ministry of Health and Welfare, was started in 1981. It involves about 2400 sentinel medical institutions (SMI) (mainly paediatric hospitals and clinics), accounting for approximately 8% of the total number of paediatric hospitals and clinics throughout the entire country. The health centre plays a great part in this system, and covers one or more cities, towns and villages. The numbers of SMI in the areas covered by health centres are approximately proportional to their population sizes. The populations covered by health centres vary widely in size from less than 30 000 to more than 250 000. SMI were recruited on a voluntary basis. Each SMI reports the numbers of cases of 16 infectious diseases (Table 1Go) to the area health centre weekly. The communication of this information among health centres, local governments (prefectures) and the Ministry of Health and Welfare is made through an on-line computer network.


View this table:
[in this window]
[in a new window]
 
Table 1 Percentiles of the numbers of cases per week per sentinal medical institution (SMI), the critical values for the onset and end of an epidemic, and the number and length of epidemics
 
Surveillance data and determination of epidemics
We used 165 604 reports from infectious disease surveillance in Japan, which covered 95% observations over 261 weeks (1993– 1997) in each of 663 health centre areas. To express the disease incidence in health centre areas, the number of cases of 16 diseases per week per SMI was used as an index.

Epidemic periods in health centre areas were determined according to their indices. We assumed that an epidemic began in the week when its index exceeded a certain value (the critical value for the onset of an epidemic) and that the epidemic ended the week before indices in four successive weeks were lower than another given value (the critical value for the end of an epidemic). The critical values for the onset of an epidemic of 16 diseases were given between 95 and 99 percentiles of the distribution of indices. The critical values for the end of an epidemic were given as 90 percentiles of the distribution. The critical values for the onset and end of an epidemic were modified if the number of epidemics greatly exceeded or undershot 1200 (1.8 per health centre in 5 years).

Table 1Go shows the percentiles of the distribution of indices, the critical values for the onset and end of an epidemic, and the number and mean length of epidemics. The number of epidemics of 16 diseases was between 800 and 1600. The mean length of epidemics was between 2 and 22 weeks.

Method for detecting early stage of epidemic
We used a simple method for detecting the early stage of an epidemic, which is that if the number of cases per week per SMI in a health centre area exceeds a critical value (critical value for an epidemic warning), an epidemic will begin in that area in the following 4 weeks (a warning for an epidemic).

This method is illustrated using a hypothetical example in Figure 1Go. Consider a warning for an epidemic of influenza-like illness in a health centre area. The index exceeds the critical value for the onset of an epidemic (given as 40) in the 12th week, and is lower than the critical value for the end of an epidemic (given as 10) from the 19th week. Thus, the epidemic period is between the 12th and 18th weeks. The pre-epidemic period (defined as the 4 weeks before the epidemic) is the 8th to 11th week. Non-epidemic periods are between the 1st and 7th weeks and between the 19th and 52nd weeks. Excluding the epidemic period, the indices in the 11th and 40th weeks exceed the critical value for an epidemic warning (given as 10). As the 11th week is in the pre-epidemic period, this warning is a true positive. Inversely, the warning in the 40th week during the non-epidemic period is a false positive.



View larger version (17K):
[in this window]
[in a new window]
 
Figure 1 A hypothetical example of epidemic and warning in influenza-like illness in a health centre area

 
The sensitivity, specificity and positive predictive value for the epidemic warnings in 1993–1997 were evaluated by setting several critical values for epidemic warnings. The critical values included one which gave the specificity of about 95%. The sensitivity was the proportion of epidemics with true positive warnings in their pre-epidemic periods. The specificity was the proportion of weeks without a warning in non-epidemic periods. The positive predictive value was the proportion of true positive warnings among all warnings.


    Results
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Tables 2 and 3GoGo show the sensitivity, specificity, positive predictive value, and the proportions of true positive warnings among the warnings 1–4 weeks before an epidemic. Increasing the critical value for an epidemic warning caused the sensitivity to decrease, and the specificity and positive predictive value to increase. The proportions of warnings in the 4 weeks before an epidemic also decreased, while that in the one week before an epidemic increased. In 10 of the 16 diseases, the sensitivity was more than 60% and the specificity was more than 95%. With these diseases, the positive predictive value was between 15.6% and 31.4%, and the proportion of warnings 3 and 4 weeks before an epidemic was between 16.0% and 68.6%. With four diseases (influenza-like illness, rubella, hand-foot-and-mouth disease, and herpangina), the sensitivity was more than 80% and the specificity was more than 95%.


View this table:
[in this window]
[in a new window]
 
Table 2 Sensitivity, specificity and positive predictive value of epidemic warnings (Part 1)
 

View this table:
[in this window]
[in a new window]
 
Table 3 Sensitivity, specificity and positive predictive value of epidemic warnings (Part 2)
 
Table 4Go shows the number of health centre areas with warnings and epidemics of influenza-like illness in all prefectures of Japan in January 1995. The prefectures are listed from the northernmost to the southernmost in the table. Among 663 health centre areas, 42 had warnings and 33 had epidemics in the first week. The number of areas with warnings increased between the first and 2nd weeks, and decreased between the 2nd and 5th weeks. The number of areas with epidemics increased between the first and 5th weeks. The weeks with peak numbers of warning and epidemic areas in this period differed among prefectures.


View this table:
[in this window]
[in a new window]
 
Table 4 Number of health centre areas with warnings and epidemics of influenza-like illness in January, 1995
 

    Discussion
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
The high specificity of epidemic warnings for the surveillance of infectious diseases is essential because the surveillance is a routine system.10.11 When the specificity of the epidemic warnings was more than 95% in the present study, the sensitivity was more than 60% in ten diseases, and more than 80% in four diseases. Thus, the early stage of an epidemic might be detectable in all these diseases, or at least in the four diseases with sensitivities of more than 80%. Among the above ten diseases, the positive predictive value was between 15.6% and 31.4%. The positive predictive value of the epidemic warnings was not expected to be high because epidemics were not frequent (1.8 times per health centre in 5 years).

Warnings 3 and 4 weeks before an epidemic might be useful for control of the epidemic. Such warnings accounted for between 16.0% and 68.6% of the true warnings in the above ten diseases. Even if a warning 1 or 2 weeks before an epidemic in an area is too late for control of the epidemic in that area, it might be useful for surrounding areas, because epidemics of many infectious diseases generally begin in small areas and subsequently spread widely. This fact is shown using an example of influenza-like illness in Table 4Go. Thus, the warnings for epidemics in small areas might provide information on epidemics in large areas.

The usefulness of epidemic warnings is related to the possibility of taking countermeasures against the epidemic. Vaccination is an efficient countermeasure for the prevention and control of epidemics of influenza-like illnesses and rubella,4,16 and for these diseases the specificity of the warnings was more than 95% and the sensitivity was more than 80%. In planning prevention and control of an epidemic, characterizing the aetiological agents, such as the type of virus, is as important as evaluating the frequency of cases.6,10 The epidemic warnings presented herein are based only on the frequency of cases. Information on characterizing the aetiological agents is obtained from laboratory-based surveillance.

There were some problems and limitations in the present study. A key problem was the definition of an epidemic, which was done by several methods and not standardized.2,5 The determination of an epidemic in the present study, described above, was according to the number of cases per week per SMI. The critical values for the onset and end of an epidemic might vary by the kind of disease, season, area and so on. In particular, we think that the definition for an epidemic warning should be based on the need for a public health response. We assumed that the pre-epidemic period was the 4 weeks before the epidemic. A pre-epidemic period of over 4 weeks could be discussed. For detecting epidemics in their early stage, the method used requires a stable surveillance system as well as other methods including time series analysis.13,15 We chose the simple method because it would be more practical for public health activities. The sensitivity, specificity and positive predictive value of epidemic warnings might be improved by using more complex methods. The epidemic warnings in the present study were on a health centre area basis because the health centre is the basic unit in the public health activities, including surveillance, against the infectious diseases in Japan. Our results might include potential biases because sentinel medical institutions were recruited on a voluntary basis. For applying the proposed method of epidemic warning to surveillance, further discussion of disease-specific issues would be important.


    Acknowledgments
 
This study was supported by a Grant-in-Aid from the Ministry of Health and Welfare, Japan, for Research on Emerging and Re-emerging Infectious Diseases.


    References
 Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 
1 Tillett HE, Spencer I. Influenza surveillance in England and Wales using routine statistics. J Hyg—Cambridge 1982;88:83–94.

2 Costagliola D, Flahault A, Galinec D et al. A routine tool for detection and assessment of epidemics of influenza-like syndromes in France. Am J Public Health 1991;81:97–99.[Abstract]

3 Sprenger MJW, Mulder PGH, Beyer WEP et al. Influenza: relation of mortality to morbidity parameters—Netherlands, 1970–1989. Int J Epidemiol 1991;20:1118–24.[Abstract]

4 Matter HC, Cloetta J, Zimmermann H et al. Measles, mumps, and rubella: monitoring in Switzerland through a sentinel network, 1986–94. J Epidemiol Community Health 1995;49(Suppl. 1):4–8.[ISI][Medline]

5 Snacken R, Lion J, Casteren V et al. Five years of sentinel surveillance of acute respiratory infection (1985–1990): the benefits of an influenza early warning system. Eur J Epidemiol 1992;8:485–90.[ISI][Medline]

6 Thacker SB, Stoup DF. Future direction for comprehensive public health surveillance and health information systems in the United States. Am J Epidemiol 1994;140:383–97.[Abstract]

7 Canada Ministry of Health. Influenza in Canada—1996–1997 season. Can Commun Dis Rep 1997;23–24:F1–F6.

8 Mar CD, Pincus D. Incidence patterns of respiratory illness in Queensland estimated from sentinel general practice. Aust Fam Physician 1995;24:625–32.[Medline]

9 Ohshiro H, Kawamoto K, Nose T. Surveillance system of infectious diseases in Japan. J Epidemiol 1996;6:S81–S85.[Medline]

10 Thacker SB, Choi K, Brachman SB. The surveillance of infectious diseases. JAMA 1983;249:1181–85.[Abstract]

11 Stroup DF, Whaton M, Kafadar K et al. Evaluation of a method for detecting aberrations in public health surveillance data. Am J Epidemiol 1986;137:373–80.[Abstract]

12 Choi K, Thacker SB. An evaluation of influenza mortality surveillance, 1962–1979. I. Time series forecasts of expected pneumonia and influenza deaths. Am J Epidemiol 1981;113:215–26.[Abstract]

13 Stroup DF, Thacker. A Bayesian approach to the detection of aberrations in public health surveillance data. Epidemiology 1993;4:435–43.[ISI][Medline]

14 Cliff AD, Haggett P, Stroup DF et al. The changing geographical coherence of measles morbidity in the United States, 1962–88. Stat Med 1992;11:1409–24.[ISI][Medline]

15 Nobre F, Stroup DF. A monitoring system to detect change in public health surveillance data. Int J Epidemiol 1994;23:408–18.[Abstract]

16 CDC. Prevention and control of influenza: recommendation of the Advisory Committee on Immunization Practices (ACIP). MMWR 1997; 46(RR-9):1–25.[Medline]