Information Technology Reference
In-Depth Information
provide interested people (practitioners, pharmacists etc.) with warnings against
approaching influenza waves. So, we have marked those time points of the former
courses where we, in retrospect, believed a warning would have been appropriate; e.g.
in the 17
th
week of the 2000/2001 season (see fig.1). This means that a solution of a
four weeks course is a binary mark, either a warning was appropriate or not.
For the decision to warn, we split the list of the most similar courses in two lists.
One list contains courses where a warning was appropriate; the second list gets the
other ones. For both of these new lists we compute their sums of the reciprocal
distances of their courses to get sums of similarities. Subsequently, the decision about
the appropriateness of a warning depends on the question which of these two sums is
bigger.
3.5
Learning
In section 3.3. we have introduced two threshold parameters X and Y. However, we
have not explained how we are getting good settings for them. In fact there is no
chance to know them. Since they are very important for the solution, namely the
decision whether to warn, we attempt to learn them. For each of our complete
influenza courses (from October to March), we have made the same experiment; we
used it as query course and we have tried to compute the desired warnings with the
remaining courses as case base. Therefor we have varied the values for the threshold
parameters X and Y. So far, we have not learnt single optimal values but intervals for
the threshold parameters. With combinations of values within these intervals all
desired warnings can be computed.
4
First Results
So far, we have developed a program that computes early warnings of approaching
influenza waves for the German federal state Mecklenburg-Western Pomerania. As we
receive data since 1997, our case base just contains four influenza periods. For each of
them, our program is able to compute the desired warnings by using the other three
periods as case base. However, the last influenza epidemic in Western Europe, where
doctors even ran out of vaccine, occurred in winter 1995-96 [20]. Unfortunately, we
do not have data for this period. Nevertheless, we hope to be able to predict such
epidemics with the help of our data of the recent influenza waves.
At present the computed warnings and follow-up warnings are only displayed on a
machine. In the near future we intend to send them by email to interested people.
Furthermore, so far we have focussed on the temporal aspect of influenza waves
for the whole federal state. Very recently, we have started to apply our program to
smaller units, namely to 6 cities or towns and to 12 districts in Mecklenburg-Western
Pomerania. Since we only receive data of written unfitness for work from the main
health insurance company, incidences for some of these units are rather small (even
the peaks are sometimes below 100 per week). For such units our program has
difficulties to determine whether an increase is already the beginning of an influenza
wave or if it occurred just accidentally. The general problem is that the smaller the
incidences are, the higher is the influence of coincidence.
