The 2011-12 influenza season was unusually late and mild, and set a new record for the season with the lowest and latest peak of influenza-like illness (ILI) since surveillance efforts began

In contrast to the 2011-12 season, the ongoing 2012-13 season is off to an unusually early and severe start, despite the fact that the national climate this past autumn was close to the seasonal average. Here we analyzed the weekly time series of confirmed influenza cases in the US from the 1997-98 influenza season to present. Our findings indicate that influenza epidemic severity and time of onset is significantly associated with the average winter temperature during the previous season, with severe and early influenza seasons being much more likely following a mild winter.

In the event of continued global warming, warmer than average winters are expected to occur more frequently, but variability in seasonal temperatures will of course remain, and average winters will still occur with regularity for some time to come. Our work suggests that mild influenza seasons during unusually warm winters are a harbinger of the likelihood of an unusually severe season to come. Hence, these findings could guide improved prevention efforts, including progressive vaccination programs after a mild winter to achieve high vaccination coverage well in advance of the next influenza season.

Time series of influenza B, A(H1N1) and A(H3N2) for annual seasons between 1997-98 to 2012-13 are shown in Figure 1. Both influenza B and A(H3N2) epidemics are off to an extraordinarily early start this season, made even more remarkable for the fact that the vaccine this season is a reasonably good match to the circulating strains, and also that early epidemic seasons in all previous years of surveillance data were due to the spread of one subtype only.

The data for each season is shown relative to Jan 1st of that season. The extraordinarily early rise in cases of influenza A(H3N2) and influenza B this season is made even more unusual by the fact that the simultaneous early epidemic rise of two strains has not hitherto been observed, and the vaccine composition is a good match to the circulating strains.

To calculate the population weighted average climate across US regions, the population and geospatial location information for all urban centers in the US with population greater than 50,000 were obtained from the US Census Bureau (comprising almost 500 population centers, and over 75% of the population of the US)

In Figure 2 we show the US population weighted average temperature during winter (January-March) and autumn (September-December) from 1997 to 2012. The winter of early 2012 was unusually warm, whereas temperature in autumn of 2012 was close to the seasonal average of the study period).

The temperature in early 2012 was significantly higher than average, whereas the temperature during autumn 2012 was consistent with the average.

For seasons where the influenza epidemic was not suppressed, we use the NCDC climate data to estimate the average temperature, “T”, over the time period comprising the initial epidemic growth phase.

For each influenza season from 1997-98 to 2012-13, we assessed the rate of exponential rise of the influenza epidemic for all strains, and total ILI mortality during the epidemic. We excluded A(H1N1) data from the 2009-10 pandemic season due to the very unusual dual wave nature of the epidemic curve (although it should be noted that the conclusions of the study remained unchanged even when those data were included).

We found the initial growth rate to be significantly correlated with ILI mortality across seasons (r=45%, p=0.04). In this analysis we thus used the initial growth rate for influenza seasons as an indicator of epidemic severity.

The initial growth rates we estimated for influenza A epidemics are on average 15% higher compared to those we estimated for influenza B epidemics (Student's t-test p=0.03), which is in agreement with results of previous studies

In Figure 3 we show the initial growth rate (r) and epidemic peak timing (t_peak) as a function of the growth rate of the season before (r_bef) and the average temperature the season before (T_bef). Significant associations are seen. In Table 1 we show the partial correlations between the growth rate and time of the peak to the other variables, after adjustment for the accuracy of the vaccine match to the circulating strain (v). We found a significant association between epidemic peak timing and epidemic growth rate when compared to the temperature and growth rate the season before, with very similar patterns seen for both influenza A and influenza B. In particular, mild winters were found to be significantly associated with early and severe epidemics the next season, even after taking into account the epidemic growth rate during that mild winter and the temperature during the next season; we found that when a winter was mild, on average 72% of the time the next epidemic was more severe than average, with epidemic growth rate 40% higher than average, and a peak timing occurring 11 days earlier than average. In addition, the relative likelihood of the following epidemic peaking before January 1^{st} was over 80% higher.

We posit that warm winters are more likely to be succeeded by early and severe influenza seasons due to fewer people being infected due to the warm weather, thereby leaving an unnaturally large fraction of susceptible individuals in the population going into the next season. The severity of the next season could potentially be exacerbated by the early onset, if the onset occurs before most of the population has had the opportunity to be vaccinated.

Our observation that climate patterns can have a profound impact on influenza epidemics beyond just the time frame of the current season will likely open up interesting avenues of further research.

Correlation of the exponential growth rate (r) and epidemic peak timing (t_peak), to the growth rate and average temperature during the season before, (r_bef) and (T_bef), respectively. The calculation of the correlations takes into account the temperature during the initial epidemic period (T) and the accuracy of the vaccine match (v) for each season under consideration. The numbers in square brackets indicate the 95% CI. The correlations in the influenza A column are the aggregated results of influenza A(H3N2) and A(H1N1) epidemics.

Spearman |
Influenza A |
Influenza B |
All Influenza |
---|---|---|---|

ρ(r,T|r_bef,T_bef,v) | -0.28[-0.38,-0.18] | -0.39[-0.50,-0.26] | -0.35[-0.43,-0.28] |

ρ(r,r_bef|T,T_bef,v) | -0.06[-0.16,0.05] | -0.11[-0.24,0.03] | -0.06[-0.15,0.02] |

ρ(t_peak,r_bef|T,T_bef,v) | -0.07[-0.19,0.06] | 0.09[-0.06,0.24] | -0.06[-0.16,0.04] |

ρ(r,T_bef|T,r_bef,v) | 0.15[0.04,0.25] | 0.28[0.15,0.40] | 0.17[0.09,0.25] |

ρ(t_peak,T_bef|T,r_bef,v) | -0.18[-0.30,-0.05] | -0.31[-0.44,-0.17] | -0.11[-0.20,-0.01] |

ST is grateful to the organizers and funding agents of the recent Workshop on Mathematical and Statistical Modeling with Applications in Epidemiology and Finance at Northeastern Illinois University; stimulating workshop discussions prompted the initiation of these studies.

This is great work you all did!!