Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 125

Figure 15. Lift function for three-variable input
records have a higher level of prediction than just chance. Therefore, we can concentrate on these 4
deciles of patients. If we use the expanded model that includes patient demographic information plus
additional diagnosis and procedure codes for COPD, we get the lift shown in Figure 16. The model can
now predict the first 5 deciles of patient outcomes.
Therefore, we can predict accurately those patients most at risk for death; we can determine which
patients can benefit from more aggressive treatment to reduce the likelihood that this outcome will oc-
cur.
PredIctIve modelIng of contInuous outcomes
We next turn our attention to predictive models using length of stay, cost, or charges as the outcome vari-
able. In this case, we are trying to determine whether patients of similar severity have similar outcomes.
Not all of the available predictive models will work with interval outcomes, so we will reduce the model
choices to neural networks, memory based reasoning, regression, and decision trees.
Because we are no longer concerned with a rare occurrence as a target, we do not have to sample
before the analysis. Therefore, we use the predictive model as given in Figure 17. It is not necessary to
use the entire sample; a subsample will give nearly the same results and reduce the computational time,
which can be considerable if the data set is extremely large. We will use different subsamples to examine
the data. We use length of stay as the target variable, and limit our analysis to pneumonia, septicemia,
and immune disorder. We also include mortality as an input variable. We start with a 1% sample. Ac-
cording to the model comparison, the optimal model is a decision tree with the smallest average error
(Figure 18).
The reporting results will be different compared to a binary model. There is no misclassification rate
reported, nor is there lift. Instead, the focus is on the average error of the model.
Figure 16. Lift function for complete model