Explaining Medical Model Exceptions - Computational Intelligence in Healthcare: Advanced Methodologies - page 273

6 Discussion
Though statistics is the traditional method for data analysis, other intelligent data
analysis methods have become popular, especially for large quantities of data, e.g. for
online analytic processing [30]. Such “data mining” methods include artificial neural
networks, Bayesian networks, decision trees, genetic algorithms, and statistical pat-
tern recognition [31]. However, ISOR is designed for medical studies. Usually, there
are just a few cases involved in such studies and the data sets are rather small.
Outlier detection and outlier management are interesting research topics, especially
in medicine, because often measuring errors occur due to nurses and doctors being
under stress. A strange data value that stands out because it is not like the rest of the
data in some sense is commonly called an outlier [32]. However, in the strongest
model version in the presented application of dialyse and fitness 22 out of 72 cases
(this means about 30 %) contradict the hypothesis (table 1). Since the threshold of
most statistical tests is 5%, in the presented application the contradicting cases should
not be treated as outliers.
Furthermore, the idea of ISOR is to support research doctors in their search for rea-
sons why cases are deviating from a research hypothesis. Usually, the doctors are not
interested in this question when there are just few outlier cases. They are becoming
curious when the number of such cases is rather big.
How about the ethical point of view? Patients without health insurance or with
serious co-morbidities can become fiscal disasters to those who care for them.
Papadimos and Marco [33] presented a philosophical discourse, with emphasis on the
writings of Immanuel Kant and G.F.W. Hegel, as to why physicians have the moral
imperative to give such “outliers” considerate and thoughtful care. However, the seri-
ously ill dialyse patients should not be blamed if they do not go in for sports actively,
because they might feel too week due to their physical condition.
Furthermore, ISOR is a general program applicable on medical studies. In the dial-
yse and fitness application the patients had the choice to actively participate in the
fitness program or to do so rather passively. Usually, in medical studies patients have
just the choice between different treatments (often a rather new one and an established
one). Beforehand it often cannot be said which choice will lead to an outlier group.
7 Conclusion
In this chapter, it has been proposed to use CBR to explain cases that do not fit a sta-
tistical model. Here one of the simplest models was presented. However, it is rela-
tively effective, because it demonstrates statistically significant dependencies. In the
example of fitness activity and health improvement of dialysis patients the model
covers about two thirds of the patients, whereas the other third can be explained by
applying CBR.
The presented method makes use of different sources of knowledge and informa-
tion, including medical experts. This approach seems to be a very promising method
to deal with a poorly structured database, with many missing data, and with situations
where cases contain different sets of attributes.