Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 130

Figure 22. Results of predictive modeling for length of stay with 1000 random observations
Figure 23. Results of regression using 1000 observations
do, however, want to add more information about patient diagnoses, using the diagnoses and procedures
listed in Tables 2 and 3 in the preceding section. We use the same sample size to examine the target value
of length of stay for patients with COPD (1% of the patients with COPD). The only change we make
to the diagram process is to change the target variable to length of stay, and to include mortality as an
input variable. We also remove rule induction as a potential model. We change the sampling method to
random rather than to stratify since there is no categorical outcome on which to base the stratification.
Figure 25 shows that Dmine regression (where the interval variables are categorized) is the best model,
and that there is considerable variability in the average error for the testing set.