Biology Reference
In-Depth Information
outbreak signals occur when the disease onset and spread is faster than the
rate of reporting for the health sources, or when the disease effect is rapid
and tightly peaked. Lognormal outbreak signals are more common: many
diseases are seen to have a lognormal distribution after time of infection to
when symptoms develop and are reported (incubation time).
Buckeridge et al. (2004) has proposed a realistic model that includes model-
ing of anthrax patterns using a plume model for dispersion as well as model-
ing incubation period and behavior of the infected population. Wallstrom et al.
(2005) present a software tool (HiFide) for modeling cryptosporidium outbreak
signals and influenza in univariate health series. Both of these models are based
on real outbreaks. STEM (Ford et al., 2006) is a software plug-in that can be used
to quickly generate outbreak signals with a variety of infection parameters,
over real geographic transportation networks. Other methods, which simulate
individual cases (Wong et al., 2002) or spatiotemporal data (Cassa et al., 2005;
Watkins et al., 2007), can be adapted to generate daily counts.
However, we again caution that such outbreak signal simulations can cur-
rently only be judged via domain knowledge; there is not enough data to
compare their accuracy using statistical tests. For this reason, our outbreak
simulator extends outbreak simulation to the multivariate case. Because an
outbreak will likely manifest in multiple related series, we must be able to
simulate an outbreak signal which occurs in each. The simulator can gener-
ate both spike and lognormal shapes, in a variety of sizes (increased num-
ber of affected cases, possibly different for different series) and shapes; it
allows flexibility to tailor the outbreak generation as appropriate for the
comparison.
In addition, we provide a novel labeling system that takes the multivar-
iate nature of the data into account. There is still debate as to what time
period of an outbreak it is valuable to detect. For example, if a detected alert
occurs after the peak of the disease effect, it is not very useful to the public
health practitioners. In general, for univariate series, the debate is between
counting any alert during an outbreak versus counting only alerts that occur
before the peak of the disease outbreak signal. For the multivariate case, this
is even more complicated, as the peaks may occur at different times in differ-
ent series. For this reason, instead of only two labels (normal/outbreak), we
use four labels as follows:
0
—no outbreak signal on that day
a
—outbreak before any series have peaked
b
—outbreak between the first and the last series' peak
c
—outbreak after all series have peaked
Note that labels are applied to days, not to single series. An example is shown
in Figure 2.3, where a single lognormal outbreak signal was placed on day
10 and injected into both series. Until day 9, the label is 0. On day 10, an
