Influenza Surveillance System in Japan

The systematic surveillance of influenza and influenzalike illness (ILI) as notifiable diseases under Infectious Disease Control Law began in 1981 in Japan. Each ILI case is defined on the basis of a sudden fever >38°C, respiratory symptoms, and myalgia. The number of patients with ILI is reported on a weekly basis from ≈2,400 sentinel pediatric and general physicians and 663 health centers throughout Japan. The number of sentinels is decided on the basis of the size of the population of the health center area where they serve: a health center with population <75,000 would have one sentinel, a population 75,000–125,000 would have two, and populations >125,000 would have [(population – 125,000)/100,000] sentinels (10). The number of sentinels per million is ≈20; recruitment is on a volunteer basis. Sentinels forward clinical data to ≈60 prefectural or municipal public health institutes, and data generated are electronically reported to the Infectious Disease Surveillance Center in the National Institute of Infectious Diseases (Tokyo) (5,6).

Geographic Analysis

We analyzed surveillance data from 1992 to 1999 for 46 prefectures, excluding Okinawa Prefecture, which is approximately 800 km from the four major islands of Japan. To combine data from all prefectures and examine trends at the national level, we calculated the number of reported ILI cases per sentinel per week after adjusting the epidemic curves with a 5-week unweighted moving average (reported ILI cases per sentinel per week [RC/S/W]) as an indicator of ILI activity. This procedure smoothed the data and simplified identification of the seasonal peak (7). The peak week during each influenza season was defined when the greatest unweighted moving average was observed in individual prefectures. For the time scale of geographic analysis, the first week was defined when the first peak was observed in any of the prefectures during the season; subsequent weeks were then numbered accordingly.

For the spatiotemporal spread of the 1992–1999 epidemics in Japan, we used the kriging geostatistical method to estimate point values by using surrounding, known point values (11,12). The address of the prefecture government was used as the representative site of prefecture surveillance data, and unweighted moving average for each prefectural peak week was applied after adjusting the time scale. Kriging uses a weighted moving average interpolation to produce the optimal spatial-linear prediction. The estimated kriging weight matrix is a product of the inverse covariance weight matrix and the distance matrix (11,12).

To make kriging maps as contour maps showing the timing of peak ILI activity, we performed the following steps. First, we created an empiric semivariogram to examine the structure of data. The empiric semivariance is 0.5 times the difference squared, when Euclidean distance is used. Second, this semivariogram estimated the theoretical model parameters through a weighted least-squares technique. The data showed a spherical pattern. Next, the weights were determined by incorporating the spherical pattern of covariance. Finally, we estimated the values at unmeasured points and made filled-contour maps from the kriging weights for the measured values. The isobars on the contour maps represent interpolated time of peak activity distributed spatially and were placed at 1-week intervals. All procedures were carried out on ArcGIS 8.2 (ESRI, Redlands, CA) and Geostatistical Analyst (ESRI) for Windows.

Statistical Analysis

To ascertain the relationship between epidemic scale and velocity of spread, we used three parameters. greatest number of ILI cases, increasing-to-peak period, and nationwide peak-duration. The first parameter was the greatest number of ILI cases, defined as the greatest number of RC/S/W in each prefecture The second parameter was the increasing-to-peak period, defined as the time from the week RC/S/W was >50% of peak to the week of the peak. Influenza epidemics usually show an elevated incidence of ILI before the peak and for some weeks after each epidemic. In our study, sharp increases in ILI cases were seen in the weeks before the epidemic; we focused on these weeks. The means for the two parameters across prefectures were calculated for each season. The third parameter was nationwide peak-duration, defined as the time between the first and last week that showed the greatest number of ILI cases among 46 prefectures in each season. Spearman’s correlation coefficient was used to analyze the relationship between all pairs of the three indexes. All calculations were performed with Microsoft Excel 2002 (Microsoft Corp., Redmond, WA), and significance was determined at p < 0.05.

Influenza Epidemics, 1992–1999

Figure 1

Figure 1. Moving averages of weekly reported influenzalike illness (ILI) cases.

We analyzed 2,586,272 ILI cases during the 7-year period from 1992 to 1999. The annual influenza season began between November and December, peaked between January and February, and returned to baseline between April and June for the study period in every year at all reporting sites (Figure 1). Seasonal peaks in ILI activity occurred annually in all prefectures. Nationwide epidemics lasted for 3–4 months, but successive or overlapping waves of infection by influenza A and B sometimes resulted in a more prolonged outbreak, as in the 1996–1997 season.

The predominant circulating strain was A/H3N2 in all seasons, except in 1996, when it was A/H1N1. Larger scale epidemics were observed in the 1992–93, 1994–95, and 1997–98 seasons, when new antigenic variants of A/H3N2 as predominant circulating subtypes were isolated, namely A/Kitakyushu/159/93, A/Wuhan/359/95, and A/Sydney/5/97, respectively. A relatively large-scale epidemic was also observed in the 1998–99 season, but the predominant circulating subtype was A/Sydney/5/97, as in the previous year. The first peak arose from one predominant viral agent and was bigger than the second peak in the 1996–97 season, with a bimodal curve.

Geographic Analysis

Figure 2

Figure 2. Timing of peak influenzalike illness epidemic activity by week in Japan. The isobars on the contour maps represent interpolated time of peak activity distributed spatially at 1-week intervals. The first week...

Kriging analysis clearly illustrated spatiotemporal movement of ILI epidemics in Japan (Figure 2). Seasonal ILI activity occurred in a sequential manner, and differences between seasons were easy to identify and characterize. The starting prefectures or areas of the peak ILI activity were mostly in the western part of Japan, except in the 1996–97 season. Trends did not change with the appearance of new variants. The most dramatic differences from year to year were in spreading pattern, as shown in the contour map of peak ILI activity by week. With the monotonous spreading pattern, peak ILI activity covered Japan within 3 to 5 weeks in large epidemics with new antigenic variants of A/H3N2, such as occurred in the 1992–93, 1994–95, 1997–98, and 1998–99 seasons. On the other hand, with the multitonous patterns, peak ILI activity covered Japan within 12 to 15 weeks in small epidemics without new antigenic variants of A/H3N2 in the other four seasons.

Statistical Analysis

Figure 3

Figure 3. Correlation analysis among three parameters. Greatest number of cases refers to the greatest number of reported influenzalike illness (ILI) cases per sentinel per week (RC/S/W) in each prefecture in each season....

During the 7-year study period, the greatest number of ILI cases, the increasing-to-peak period, and the nationwide peak duration were 4.67–40.88 ILI cases per sentinel per week, 3.43–4.83 weeks, and 3–15 weeks, respectively (Figure 3). With the larger epidemics, such as in 1992–93, 1994–95, and1997–98, and, to a lesser extent, 1998–99, the greatest number of ILI cases was >28 RC/S/W, the increasing-to-peak period was <4 weeks, and the nationwide peak duration was <5 weeks. These three parameters were interrelated (p < 0.05).