Explaining Medical Model Exceptions - Computational Intelligence in Healthcare: Advanced Methodologies - page 271

To test the method a random set of parameter values was deleted from the observed
data set. Subsequently, the method was applied and it was attempted to impute the
deleted values - but not the ones actually missing!
Table summarises how many deleted values could be imputed. Since for those 12
parameters that were only measured once and remain constant no values were deleted
(and none of them are really missing), they are not considered in table 4. More than
half of the deleted values could be at least partly imputed, nearly a third of the deleted
values could be completely imputed, about 58% of imputation occurred automati-
cally. However, 39% of the deleted values could not be imputed at all. The main rea-
sons are that for some parameters no proper method is available and that specific ad-
ditional parameter values are required that are sometimes also missing.
Another question concerns the quality of the imputation. That means how close
are the imputed values to the real values? It has to be differentiated between exact,
estimated and binary imputed values. Just one of the 13 binary imputed values was
wrong. However, this mainly shows the “quality” of the expert user, who probably
was rather cautious and made binary assessments only when he/she felt very sure.

Table 4. Summary of randomly deleted and imputed values. Only the deleted values were at-
tempted to impute, but not the really missing ones.

Number of Parameters 112

Number of values 448

Number of really missing values 104

Number of randomly deleted values 97

Number of completely imputed values 29

Number of estimated values 17

Number of partly imputed values (binary) 13

Number of automatically imputed values 34

Number of expert assistance 25

Number of values that could not be imputed 38

Table 5. Closeness of the imputed values. The numbers in brackets show the deviations on
average in percentage.

Deviation Number of exactly imputed

values

Number of estimated values

< 3 % 14 (2.2) 9 (1.8)

< 5 % 13 (5.7) 5 (6.1)

< 10 % 2 (8.5) 2 (7.4)

> 10 % 0 1 (12.3)

The deviation (percentage) between the imputed values and the real ones is shown
in table 5. Concerning the two exactly imputed values with more than 5% deviation,
we consulted the expert user, who consequently altered one formula, which had been
applied for both values. For the estimated values, it is conspicuous that for few values