Information Technology Reference
In-Depth Information
comparison to the charlson Index
We first look at the relationship of the Charlson Index to the disease staging: length of stay (Table 5).
Note that the Charlson code of 0 is divided between the first three disease staging levels with a
Charlson code of 1 divided between 50% in the disease staging level 3 and 25% in level 4. By Charlson
index 4, the majority of patients are contained within disease staging level 4. There is only a slight shift
to index 5 at Charlson index level 11 and above.
Figures 1,2, and 3 compares the Disease Staging: Resource Level (DS) to the individual conditions
that define the Charlson Index using a series of pie charts. Virtually every condition is defined using
DS levels 3, 4, and 5. What does change is the proportion of levels 3 and 4. For both acute myocardial
infarction and congestive heart failure, half of the patients have staging level 3 with a small proportion
in the most severe category of 5. There is a larger proportion of level 3 for diabetes and liver disease
with a higher proportion of level 4 for metastatiic cancer and paraplegia. From this spread, it is clear
that there is no one disease that is considered in the most severe category in and of itself.
Since every one of these conditions is high on the disease staging level, this result explains why only a
value of zero for the Charlson Index is in the first three disease staging levels. It shows that the Charlson
Index is focused on very severe patients compared to the resource demand index, which appears to focus
on more mild conditions and has a much higher proportion of patients in the more severe categories.
Similarly to the predictive model using the Charlson Index to predict the APRDRG index, we can use
a predictive model to determine whether the Charlson Index can compare to the disease staging level.
Figure 4 gives the variable definitions and identified roles in the predictive model; Figure 5 gives the
predictive model. The misclassification rate is given in Figure 6, and the Lift is given in Figure 7.
Note that the misclassification rate is much higher compared to the predictive model for the APRDRG
in the previous chapter. The decision tree remains the best model as it was in Chapter 5.
However, the lift results indicate that the model is effective only to the second decile, indicating that the
Charlson Index can make very few predictions accurately of the disease staging: resource demand level.
Figure 8 has part of the decision tree; only a small portion of this tree is dependent upon the Charlson
Index after total charges and age at admission. Therefore, the relationship between the Charlson Index
and the disease staging model is very slight.
If the Charlson Index is 0, the next measures are for total charges and admission type. If the Charlson
Index is greater than one but less than or equal to 3, then admission type is next in the tree with a slight
division for resource level 4. Otherwise, the admission type then splits the patients into resource level
4 versus resource level 3.
It seems quite reasonable that the resource demand level would depend more upon charges and length
of stay than patient condition. Therefore, we will also consider the disease staging: mortality level as
well. We perform a similar analysis, changing only the target variable. Figure 9 gives the misclassifica-
tion rate; Figure 8 gives the Lift function.
Figure 10 indicates that the optimal model is a decision tree. However, the misclassification rate is
extremely high at 42%. The relationship between the Charlson Index and the disease staging indices is
extremely poor.
The prediction decile is now at 3 so that 30% of the patients can be classified more accurately using
the model compared to random chance. Therefore, for the most critical patients, the model has reason-
able predictability.
Search WWH ::
Custom Search