Explaining Medical Model Exceptions - Computational Intelligence in Healthcare: Advanced Methodologies - page 261

the doctor's knowledge can not be applied, but it is sensible to save it anyway. In es-
sence, there are two different ways to do this. The first one is to generate a correspon-
dent rule and to insert it into ISOR's algorithms. Unfortunately, this is very complicated,
especially to find an appropriate way for inserting such a rule. The better alternative is
to create an artificial case. Instead of the name of a patient an artificial case number is
generated. The other attributes are either inherited from the query case or declared as
missing. The retrieval attributes are inherited. This can be done by a short dialogue (fig-
ure 3) and ISOR's algorithms remain intact. Artificial cases can be treated in the same
way as real cases: they can be revised, deleted, generalised, and so on.
2.2.2 Solving the Problem “Why Did Some Patients Conditions Became
Worse?”
A set of explanations of different origin and different nature is obtained. There are
three categories of explanations: additional factor, model failure, and wrong data.
Additional factor. The most important and most frequent solution is the influence of
an additional factor. However three main factors are obviously not enough to describe
all medical cases. Unfortunately, for different patients different additional factors are
important. When ISOR has discovered an additional factor as explanation for an ex-
ceptional case, the factor has to be confirmed by the medical expert before it can be
accepted as a solution. One of these factors is Parathyroid Hormone (PTH). An in-
creased PTH level sometimes can explain a worsened condition of a patient [24]. PTH
is a significant factor, but unfortunately it was measured for only some patients.
Another additional factor as a solution is phosphorus blood level. The principle of
artificial cases was used to introduce the factor phosphorus as a new solution. One
patient's record contained many missing data. The retrieved solution meant high PTH,
but PTH data in the current patient's record was missing too. The expert proposed an
increased phosphorus level as a possible solution. Since data about phosphorus data
was also missing, an artificial case was created, that inherited all retrieval attributes of
the query case, whereas the other attributes were recorded as missing. According to
the expert, high phosphorus can provide an explanation. Therefore it is accepted as an
artificial solution or a solution of an artificial case.
Some exceptions can be explained by indirect indications, which can be considered
as another sort of additional factor. One of them is a very long time of dialysis (more
than 60 months) before a patient began with the fitness program.
Model failure. We regard two types of model failures. One of them is deliberately
neglected data. As a compromise we just considered data of six months, whereas fur-
ther data of a patient might be important. In fact, three patients did not show an im-
provement in the considered six months but in the following six months. So, they
were wrongly classified and should really belong to the “better” category. The second
type of model failure is based on the fact that the two-category model was not precise
enough. Some exceptions could be explained by a tiny and not really significant
change in one of the main factors.
Wrong data are usually due to a technical mistake or to data not really proved. One
patient, for example, was reported as actively participating in the fitness program but
really was not.