Computational quantitative projections for H1N1 flu
dynamics in the United States
Please note: The research covered on
this page is on computational forecasts of the initial
outbreak of H1N1 in the United States in the spring 2009
when case counts were still comparatively small. Our
simulations were designed to specifically target the
initial outbreak time window of the pandemic. Here, we do
not report on projections for the time course of H1N1
after June 2009. This page is no longer updated.
Last updated:
June 3, 2009 9:00pm (CDT)
On May 15, a CDC official estimated the total number of
influenza-like illnesses in the US to be beyond 100,000
with a large fraction caused by the new strain of H1N1
Influenza. This is in evident conflict with our first
estimate of 1,700 cases in the US by the end of May (green
line in the plot below). We would like to emphasize that
this figure was obtained within 4 days of the first alert
with very preliminary information. Nevertheless, as noted
in this site, our projections were subject to constant
verification and calibration. The subsequent two weeks, we
produced new projections on a frequent basis, improving the
quality of our results. In fact, within a week, our
projections for the end of May estimated 40,000 in the US
(magenta line) and by May 9 (two weeks after the
initial alert emitted by the CDC) the estimate was close to
90,000 cases in the US (red line).
As the epidemic developed and entered a sustained
and widespread phase, the official number of confirmed
cases by the CDC represent only a fraction of the actual
cases occurring.
About this Project
Initial Phase:
Projections
On April 24, WHO emitted an epidemic alert about an
influenza-like illness in the United States and Mexico.
Within a week, the situation evolved rapidly, with 11
countries reporting cases of influenza A(H1N1) infection.
Particularly at the onset of an emerging epidemic,
early projections are of vital importance to asses the
short term impact of the novel infection.
On April 28, we posted our first quantitative projections
for H1N1 flu dynamics in the US. Due to the novelty of the
disease, information about the virus was still preliminary
and subject to further investigation, therefore we
restricted our detailed projections to a four-week
timeframe.
With our projections, we correctly identified hot spots and
the overall geographical pattern of the novel virus in the
United States. We proportionately estimated the number of
cases that would be detected for each county of the
continental US.
As of May 19, H1N1 influenza was widespread in the United
States. The total number of cases was at 5,469 individuals
and the infection had been reported in all but two states.
Since the large majority of cases showed mild symptoms and
testing shifted focus to the more severe cases, the
official number of cases is no longer a good indicator for
the actual cases of illness.
Sustained Phase: Analysis of Long-Term
Impact
Since May 21, we are shifting our focus towards a detailed
analysis of the geographical pattern of the novel infection
and assessment of the long-term impacts of this new virus.
From the information obtained so far, we are also
extracting information about this virus that will allow for
a better understanding of the disease and more accurate
models.
Last Projection
The following maps show the
projected number of cases in the United States at the
county level. The range of numbers represent the most
likely outcome according to our model, a confidence
interval of 90% (see below for
explanation). This is a
worst case scenario, in which no
containment measures are taken to mitigate the spread.
We used the confirmed cases as of May 6th for
calibration and project from there on.
Because we adjust our simulations every 24–48 hours, our
projections are subject to changes. At the moment our model
projects a probable range of 6600–7900 by May 17, 2009 in
the United States.
Projection
for May 17 |
|
Video:
Projection until May 22 |
|
Java Applet:
Projection Browser |
|
|
|
The following figure indicates the time course of
the projected number of infected individuals in selected
metropolitan areas as a function of time in the next three
weeks. The vertical extend indicates the degree of
uncertainty, the 90% confidence interval.
What are these hot spots and areas? For a
few areas in the US (e.g. the Dallas area) we project the
total number of cases. We consider a larger urban area and
surrounding counties as long they lie within a typical
commuting distance of not more than ~125 miles of the
hotspot (e.g. Dallas).
What is the worst case scenario?
The worst-case scenario presented here is the
worst-case for our
model, which makes many
assumptions about the way a disease spreads and the
way people move. We attempt to make our
predictions fit real-world numbers as handsomely as
possible by making realistic assumptions, but since
the model cannot ultimately account for every
real-world factor the scenario presented here may be
better or worse than the actual outbreak.
What do the numbers in our
projections mean? Our simulations estimate the
range of expected cases for different regions in the
United States. These numbers are a
high-likelihood range, and the actual outbreak may
deviate from the projections. Deviations are
particularly likely during the onset of the outbreak
when only a very small number of cases have been
reported.
How can the worst case scenario projection be smaller
than the true reported cases? In certain areas we
project a smaller number of cases than are actually
reported. This is because the projections are
probabilistic, like the weather report. Although we
believe the projected numbers are quite likely and are
elated when the data match our predictions, larger or
smaller values can occur. As scientists, our job is
to be as candid as possible with the public and avoid
futzing with the data.
We are using high performance computational techniques and
multi-layer, large scale computer simulations to project
the time course of the H1N1 flu epidemic in the United
States. Our simulations yield projections and risk
assessments of the epidemic outbreak in a
worst
case scenario, in which no containment measures
are taken to mitigate the spread. Therefore, the
actual case numbers are expected to be
smaller as mitigation strategies and containment
efforts become effective. We are constantly updating our
forecast, taking into account new information on confirmed
cases and more precise information on the transmissibilty
and disease-specific parameters.
Our modeling is based on the current knowledge of the
disease parameters and takes into account the backbone of
spatial spread: A precise estimate of human mobility on
spatial scales between a few and a few thousand kilometers.
Our projections resolve the expected dynamics down to the
county scale (3,109 counties in mainland United States).
Details of our modeling approach are provided on the
Methodology page.