Conceived and designed the experiments: JBW. Performed the experiments: JBW. Analyzed the data: JBW. Contributed reagents/materials/analysis tools: JBW ENN. Wrote the paper: JBW ENN.
The authors have declared that no competing interests exist.
In temperate regions, influenza epidemics occur annually with the highest activity occurring during the winter months. While seasonal dynamics of the influenza virus, such as time of onset and circulating strains, are well documented by the Centers for Disease Control and Prevention Influenza Surveillance System, an accurate prediction of timing, magnitude, and composition of circulating strains of seasonal influenza remains elusive. To facilitate public health preparedness for seasonal influenza and to obtain better insights into the spatiotemporal behavior of emerging strains, it is important to develop measurable characteristics of seasonal oscillation and to quantify the relationships between those parameters on a spatial scale. The objectives of our research were to examine the seasonality of influenza on a national and state level as well as the relationship between peak timing and intensity of influenza in the United States older adult population.
A total of 248,889 hospitalization records were extracted from the Centers for Medicare and Medicaid Services for the influenza seasons 1991–2004. Harmonic regression models were used to quantify the peak timing and absolute intensity for each of the 48 contiguous states and Washington, DC. We found that individual influenza seasons showed spatial synchrony with consistent late or early timing occurring across all 48 states during each influenza season in comparison to the overall average. On a national level, seasons that had an earlier peak also had higher rates of influenza (rs = −0.5). We demonstrated a spatial trend in peak timing of influenza; western states such as Nevada, Utah, and California peaked earlier and New England States such as Rhode Island, Maine, and New Hampshire peaked later.
Our findings suggest that a systematic description of influenza seasonal patterns is a valuable tool for disease surveillance and can facilitate strategies for prevention of severe disease in the vulnerable, older adult population.
In a temperate region such as the United States, influenza epidemics occur every year with the highest activity occurring during the winter months
The dynamics of influenza outbreaks can be characterized by periodicity, severity, and a number of other parameters useful to describing the onset, duration, and intensity of outbreaks. Peak timing is one of the essential characteristics of outbreak dynamics and it reflects the time when an outbreak reaches its maximum intensity. Although it does not directly indicate the onset of disease in a population, it reflects the primary characteristic of outbreak dynamics, and therefore contains important epidemiological information. Considering the route of human-to-human transmission of influenza, the change in peak timing in adjacent geographical areas can manifest in a form of traveling waves, a special case of spatial disease dynamics. The explicit geographical transition can be distorted by less-defined structures of mass migration. Human movement (such as by air travel) and population size have been shown to impact the timing of influenza epidemics across the United States
Another characteristic, intensity, or the maximum seasonal incidence of influenza, is an indicator of disease severity. Seasonally, influenza varies significantly in its magnitude due to the strain virulence and host susceptibility. A severity index has been used to measure excess mortality during epidemic and non-epidemic influenza season. This severity index utilized cyclical regression to measure the intensity of baseline influenza. The researchers examining excess mortality also observed variability in intensity in relation to changes in circulating strains of influenza
While each parameter can contribute a significant amount of information about individual influenza seasons, it is the unique combination of these parameters that can improve predictability of seasonal influenza epidemics. Previous research has found that seasons with high intensity or disease burden are often followed by seasons with lower intensity
As influenza follows a specific pattern of seasonal variation, seasonal oscillations can be measured by a variety of techniques
Due to the significance of understanding spatiotemporal trends of influenza in the United States older adult population, our objectives were as follows: 1) to examine national trends of seasonality as well as the relationship between peak timing and intensity of influenza in the US older adult population, and 2) to assess seasonal variation across individual influenza seasons on a state-by-state basis and the relationship between peak timing and intensity of influenza seasons on the state level. To achieve these objectives we utilized a comprehensive source of hospitalization data, covering 98% of the United States older adult population
Data on hospitalization rates for influenza among older adults were abstracted from Centers for Medicare and Medicaid Services (CMS) for the years 1991 through 2004. Of the 136.2 million hospitalization records in CMS, 14.3 million (10.5%) were identified as hospitalizations for pneumonia and influenza, International Classification of Diseases, Ninth Revision, Clinical Modification Codes (ICD-9CM) 480–487. Of these, 248,889 influenza records (1.7% of pneumonia and influenza records) were analyzed. Variables available in the CMS dataset include age, gender, state of residence, date of admission, date of discharge, and up to 10 diagnostic fields. Influenza was defined using ICD-9CM 487 for any of the 10 diagnosis coding slots. Using the date of admission, we created a set of time series of weekly counts of influenza-related hospitalizations for each state and nationwide. To calculate weekly hospitalization rates, we obtained US population estimates from the US Census 1990 and 2000 and interpolated population estimates for each week of analysis.
The seasonal patterns of influenza-related hospitalization records were analyzed with a harmonic regression model adapted for a Poisson-distribution outcome
Description | Notation | Expression/Commentary |
Time series of disease rates | Per 1,000,000 | |
Time | ||
Length of time series | ||
Length of one cycle | ||
Regression parameters for intercept, sin, and cosin components | β0 β1 β2 | |
Phase shift - distance of peak from beginning of series expressed in radians | ψ | −arctan{β1/β2} |
Amplitude | (β12+β22)1/2, if β2>0; | |
−(β12+β22)1/2, if β2<0. | ||
Standard deviations for the estimates of regression parameters β1 and β2, and covariance | σβ1σβ2σβ1β2 | |
Predicted seasonal curve | ||
Peak Timing (expressed in weeks)Confidence Interval | ||
Absolute Intensity(Confidence Interval) | exp{β0+ |
|
Seasonal peak - maximum value | Smax | exp{ β0+ |
Seasonal nadir - minimum value | Smin | exp{ β0− |
To obtain the national estimate of peak timing and absolute intensity we apply the model to the whole length of the time series (596 weeks) by fitting a unique curve expressed as:
Next, we estimated peak timing and absolute intensity for each of the 13 seasons by utilizing Model 2:
Finally, for each state and each season, influenza epidemics are fitted to a unique curve expressed as:
We utilized Spearman correlations to quantify the degree of association between peak week and absolute intensity. Using Model 2, we assessed the relationship between peak week and absolute intensity for individual seasons and Model 3 allowed us to assess this association on the state level. We also examined correlations between state centroids and seasonality characteristics.
To demonstrate a spatial pattern in peak timing we compiled the results of Model 3 as a set of 13 panels (
Mapping was performed on the state level using ESRI's ArcMap GIS software (ESRI, Redlands, WA). The average peak week values highlighted in the 13 panels were categorized according to a natural breaks classification scheme with six classes.
All analysis was conducted in SAS, version 9.1 (SAS Institute) and figures were created using SPSS version 15.0.
A 13-year weekly time series of influenza hospitalization rate per 1 million older adults has regular, well pronounced seasonal curves with the highest incidence of influenza taking place in the winter months (
By superimposing a weekly time series of rates for each of the 13-seasons individually, the variability in annual weekly intensity and peak timing is evident (
MODEL 2 | MODEL 3 | CIRCULATING STRAINS (WHO) | |||||||
Season | Year | Peak Week (CI) | Intensity (CI) | Peak Week (CI) | Intensity (CI) | rs | H3N2 Strain | H1N1 Strain | B Strain |
1 | 1991–1992 | 25.76 (25.30, 26.23) | 42.51 (32.38, 55.81) | 26.38 (25.96, 26.81) | 59.47 (45.58, 73.36) | −0.46 | Beijing/353/89 | Singapore/6/86 | Panama/45/90 |
2 | 1992–1993 | 33.24 (32.45, 34.05 | 20.30 (16.03, 25.70) | 33.46 (32.93, 34.00) | 30.04 (23.03, 37.05) | 0.08 | Beijing/353/89 | Singapore/6/86 | Panama/45/90 |
3 | 1993–1994 | 27.05 (26.66, 27.45) | 51.97 (38.23, 70.65) | 27.31 (26.93, 27.69) | 75.15 (56.79, 93.51) | −0.36 | Beijing/32/92 | Singapore/6/86 | Panama/45/90 |
4 | 1994–1995 | 32.50 (31.48, 33.53) | 13.70 (10.76, 17.45) | 31.99 (31.13, 32.84) | 18.04 (14.12, 21.97) | 0.36 | Shangdong/9/93 | Singapore/6/86 | Panama/45/90 |
5 | 1995–1996 | 28.16 (27.22, 29.10) | 15.57 (12.26, 19.77) | 28.25 (27.64, 28.85) | 23.20 (17.00, 29.41) | −0.37 | Johannesburg/33/94 | Singapore/6/86 | Beijing/184/93 |
6 | 1996–1997 | 25.75 (25.16, 26.33) | 31.12 (24.34, 39.78) | 25.68 (25.25, 26.11) | 44.25 (35.57, 52.93) | 0.09 | Wuhan/359/95 | Singapore/6/86 | Beijing/184/93 |
7 | 1997–1998 | 29.03 (28.62, 29.45) | 50.61 (38.20, 67.06) | 29.02 (28.64, 29.41) | 91.62 (60.38, 122.85) | 0.41 | Wuhan/359/95 | Bayern/7/95 | Beijing/184/93 |
8 | 1998–1999 | 32.51 (32.08, 32.94) | 48.35 (37.03, 63.14) | 32.16 (31.82, 32.50) | 78.27 (57.21, 99.33) | 0.61 | Sydney/5/97 | Beijing/262/95 | Beijing/184/93 |
9 | 1999–2000 | 26.16 (25.90, 26.42) | 97.56 (69.33, 137.29) | 26.19 (25.91, 26.47) | 158.00 (112.07, 203.94) | −0.29 | Sydney/5/97 | Beijing/262/95 | Beijing/184/93 |
10 | 2000–2001 | 29.02 (27.65, 30.39) | 8.96 (6.95, 11.54) | 29.23 (28.69, 29.78) | 12.80 (9.76, 15.83) | 0.23 | Moscow/10/99 | New Caldonia20/99 | Beijing/184/93 |
11 | 2001–2002 | 32.63 (32.05, 33.22) | 29.47 (22.54, 38.54) | 32.31 (31.87, 32.75) | 49.23 (31.86, 66.61) | 0.44 | Moscow/10/99 | New Caldonia20/99 | Sichuan/379/99 |
12 | 2002–2003 | 30.70 (28.88, 32.53) | 6.31 (4.92, 8.08) | 30.37 (29.13, 31.60) | 10.67 (7.10, 14.25) | 0.12 | Moscow/10/99 | New Caldonia20/99 | Sichuan/379/99 |
13 | 2003–2004 | 23.74 (23.51, 23.98) | 116.13 (83.46, 161.60) | 23.58 (23.23, 23.93) | 157.74 (129.29, 186.19) | −0.23 | Moscow/10/99 | New Caldonia20/99 | Hong Kong/330/2001 |
Overall, absolute intensity and peak timing had a strong, inverse relationship (rs = −0.5, p<0.05); the earlier the peak in an influenza season the more intensely the season is experienced (
We estimated seasonality characteristics at the state levels using Model 3 (see supplemental material,
A dynamic movie (not included) shows the 13 seasons individually to examine synchrony with the average peak of disease. Screen shots of 3 seasons from the movie are shown in
Panel A: 1991/1992, Panel B: 1992/1993, Panel C: 1993/1994.
The estimates for intensity are highly correlated (rs = 0.97) but differ substantially indicating the regional variation in hospitalization rates (
We examined the national trend of influenza seasonality and the relationship between peak timing and intensity of influenza in the US older adult population and demonstrated a clear pattern in the appearance of seasonal peaks. The range of the average peak week shows that within a typical influenza season, the continental states exhibit a peak within 4 weeks with a general west to eastward spread; Nevada, Utah, and California were the first states in terms of peak timing to experience influenza while, on average, Rhode Island, New Hampshire, and Maine were the last. The early influenza seasons are typically more pronounced. These findings have a strong, practical application and are well supported by published research.
Previous research conducted to examine spatial trends of influenza demonstrated a similar spatial pattern. A study conducted by Grais, et al investigating the role of air travel in the forecasting of influenza showed similar movement patterns; researchers found a general west to east movement of influenza through the United States with some variability between influenza seasons. Focusing on the large air traffic hubs; Detroit, Los Angeles, Miami, New York City, and Philadelphia; it was demonstrated that influenza peaked first in Los Angeles in early December and ended in New York City in late February
We found that individual influenza seasons show synchrony with consistent late or early timing occurring across all 48 states during each influenza season in comparison to the average across the 13 influenza seasons. In a study by Brownstein et al. measuring the rate of inter-regional spread and timing of influenza in the United States for nine influenza seasons, researchers found that influenza took 2 weeks to peak over all United States regions. Using nine broad United States Census Bureau defined regions, the researchers observed similarities between the spread of influenza between seasons; specifically, geographical patterns were synchronized within individual seasons, similar to findings presented in this report. Variation was witnessed only in the week at which influenza peaked between each season. However, the authors commented that limited data available for subgroup analysis, such as by age group or on a regional level, may be overlooking detailed patterns in the spread and timing of influenza
In the present study, we aimed to examine the relationship between peak timing and intensity of influenza seasons in order to establish a base for predicting influenza. We found that across all 13 influenza seasons, peak timing and absolute intensity were significantly, inversely related; earlier seasons are more likely to have more cases of influenza on both the regional and national levels. While the inverse correlation is primarily driven by the 1999–2000 and 2003–2004 influenza seasons, it is through these outlying seasons that important public health implications can be made. For example, the 2003–2004 influenza season experienced higher than usual intensity due to a vaccine mismatch. The higher than average severity in 1999 has been hypothesized to be linked to the particularly severe 1997–1998 influenza seasons in the swine population
In the present study, we estimated peak timing using the δ-method
Incomplete testing for influenza in hospitalization records and the lack of control for circulating influenza strains in our models limit our inference to causal pathogenicity, although we have made an attempt to consider the role of strain in
Season | H3N2 Strain (WHO) | H3N2 Strain (CDC) |
H1N1 Strain (WHO) | H1N1 Strain (CDC) | B Strain (WHO) | B Strain (CDC) |
1991–1992 | Beijing/353/89 |
Beijing/353/89 |
Singapore/6/86 | Taiwan/01/86 | Panama/45/90 |
Panama/45/90 |
1992–1993 | Beijing/353/89 | Beijing/32/92Shangdong/9/93 | Singapore/6/86 | Taiwan/01/86Texas/36/91 | Panama/45/90 |
Panama/45/90 |
1993–1994 | Beijing/32/92 |
Beijing/32/92 |
Singapore/6/86 | Taiwan/01/86Texas/36/91 | Panama/45/90 |
Panama/45/90 |
1994–1995 | Shangdong/9/93 |
Shangdong/9/93 |
Singapore/6/86 | Taiwan/01/86 (50%)Texas/36/91 (50%) | Panama/45/90 |
Panama/45/90 |
1995–1996 | Johannesburg/33/94 |
Johannesburg/33/94 |
Singapore/6/86 | Taiwan/01/86 (50%)Texas/36/91 (50%) | Beijing/184/93 |
Beijing/184/93 |
1996–1997 | Wuhan/359/95 |
Wuhan/359/95 |
Singapore/6/86 | Bayern/7/95 | Beijing/184/93 |
Beijing/184/93 |
1997–1998 | Wuhan/359/95 |
Wuhan/359/95 |
Bayern/7/95 |
Bayern/7/95 |
Beijing/184/93 |
Beijing/184/93 |
1998–1999 | Sydney/5/97 |
Sydney/5/97 |
Beijing/262/95 |
Beijing/262/95 |
Beijing/184/93 |
Beijing/184/93 |
1999–2000 | Sydney/5/97 |
Sydney/5/97 |
Beijing/262/95 |
Beijing/262/95 |
Beijing/184/93 |
Beijing/184/93 |
2000–2001 | Moscow/10/99 | Panama/2007/99 | New Caladonia/20/99 |
New Caladonia/20/99 |
Beijing/184/93 | Hong Kong/330/2001 |
2001–2002 | Moscow/10/99 | Panama/2007/99 (100%) | New Caladonia/20/99 |
New Caladonia/20/99 |
Sichuan/379/99 | Yamagata/16/88 (23%) |
2002–2003 | Moscow/10/99 | Panama/2007/99 (93%) | New Caladonia/20/99 |
New Caladonia/20/99 |
Sichuan/379/99 | Hong Kong/330/2001 (99%)Yamagata/16/88 (1%) |
2003–2004 | Moscow/10/99 | Panama/2007/99 (11%)Fujian/411/2002 (89%) | New Caladonia/20/99 |
New Caladonia/20/99 |
Hong Kong/330/2001 | Yamagata/16/88 (93%)Victoria/2/87 (7%) |
*Matching WHO and CDC strains for each of the corresponding seasons and influenza strains.
**Percentages (in parentheses) represent the specimens tested positive for each of the circulating strains.
Yearly parameters captured by the proposed approach allow for analysis of complex, non-linear trends over a long time frame, examination of characteristics of individual seasons, and assessment of inter-season heterogeneity and intra-season correlations. Using complex modeling techniques, researchers can determine vaccination practices for vulnerable populations and state level clinical interventions can be enhanced by forecasting the expected intensity of disease and the timing of the disease's peak in specific subpopulations and geographical locations. Understanding the geographical patterns of influenza spread and utilizing multiple parameters for predictive modeling are essential for guiding prevention efforts.
Compiled the results of Model 3 as a set of 13 panels. Each panel depicts the 48 states in ascending order of the average peak week of the 13 influenza seasons.
(1.37 MB DOC)
Seasonality characteristics at the state level using Model 3.
(0.29 MB DOC)
We would like to thank the Centers for Medicare and Medicaid Services for providing us with the hospitalization data and our funding source, the National Institute of Allergy and Infectious Diseases (N01 AI-50032: HHSN266200500032). Many thanks to the reviewers for their thoughtful suggestions and comments. We would also like to thank InForMID research staff members for their assistance with manuscript preparation and editing.