The Charlson Comorbidity Index - Text Mining Techniques for Healthcare Provider Quality Determination - page 174

The definition of a cutpoint just serves to mask the problem of assigning weights that are too high
for specific diseases. If we consider the entire distribution, there is considerable overlap between the
subgroups in the population by the Charlson Code. Figure 11 gives the probability functions for length
of stay and total charges for a cutpoint of 2; Figure 12 gives them for a cutpoint of 5.
Note that while the peak for length of stay is at 2 days for code zero (and $10,000 for total charges),
it shifts to 4 days and $11,000 for code one. However, almost the entire curve for cutpoint one is in the
area of the cutpoint zero, suggesting that there will be little predictive ability when a cutpoint is used.
Similarly, the area for code one is in the area for code zero, suggesting that there will be little predic-
tive ability except at the high end of the scale beyond day 20 and $50,000 in cost.
cHange In dIseases to defIne neW Index
We want to suggest an alternative set of diseases to create a new index. One of the biggest problems with
such a list of diseases is what should be included, and what should be excluded. We can consider the
patient conditions most associated with mortality (Table 17). In order to find these patient conditions,
we need to investigate all fifteen columns of diagnosis codes that are available in the NIS. We limit the
codes to the first three digits. The following SAS code will give us the necessary values.
data nis.died;
set nis.sort;
where died= 1 ;
Figure 11. Probability density of length of stay and total charges for a cutpoint at 2
Figure 12. Probability density of length of stay and total charges for a cutpoint at 5