Explaining Medical Model Exceptions - Computational Intelligence in Healthcare: Advanced Methodologies - page 272

the deviation is rather big. The probable reason is that general estimation methods
have problems with varying parameter courses.

5.2 Imputation of Real Missing Data and Setting up a New Model

As a consequence of the experimental results, we assumed that our method is not per-
fect but at least practical. So, it was attempted to impute real missing data. The result
is shown in table 6.
It is no surprise that more missing values could be imputed (table 6) than randomly
deleted ones (see table 2), because all imputation methods rely on other parameter
values and in general holds that with an increased number of given values the chance
of imputing missing values increases too. In the experiment (table 4) not just the ran-
domly deleted values were missing but also the real missing ones.

Table 6. Summary of missing and imputed values.

Number of Parameters 112

Number of values 448

Number of missing values 104

Number of completely imputed values 37

Number of estimated values 24

Number of partly imputed values (binary) 19

Number of automatically imputed values 49

Number of expert assistance 31

Number of values that could not be imputed 24

After this imputation we return to the original problem, namely to set up a model
for the hypothesis that actively participating in the fitness program improves the con-
ditions of the patients (see section 2.1). Since many missing values have been im-
puted, the expert user can chose other main factors to set up the model, namely also
those ones where many data had been missing before. In fact, the expert user now
chose a different third factor than before, namely PTH instead of WorkJ. The result-
ing strongest model is shown in table 7.

Table 7. Results of Fisher's exact test, for p < 0.0001.

Patient's
physical condition

Fisher
Exact p

Active Non-active

Better 39 1

< 0.0001

Worse 11 21

The result is obviously much better than the model before (see table 1 in section
2.1). Since the missing data problem is not responsible for all exceptional cases, also
for this model some (eleven) exceptional cases still have to be explained.