The Charlson Comorbidity Index - Text Mining Techniques for Healthcare Provider Quality Determination - page 167

Table 9. Predicted versus actual mortality with a 50/50 split in the data

Predicted Value Actual Value Count Percentage

0 0 23393 23.2%

0 1 6747 6.6%

1 0 26751 26.7%

1 1 43398 43.3%

Table 10. Predicted versus actual mortality with a 75/25 split in the data

Predicted Value Actual Value Count Percentage

0 0 190073 71.1%

0 1 51922 19.4%

1 0 10507 3.9%

1 1 14937 5.6%

Table 11. Predicted versus actual mortality with a 90/10 split in the data

Predicted Value Actual Value Count Percentage

0 0 450265 89.8%

0 1 49339 9.8%

1 0 1039 0.2%

1 1 806 0.16%

diction of length of hospital stay, use of the Charlson Index in a linear regression yields a 3% r 2 value
(even though all of the variables are statistically significant).
However, in studies with small samples and small levels of mortality, the correlation to the Charlson
Index is not necessarily statistically significant.(Gettman et al., 2003) Other indices can be substituted
if they increase the level of the R 2 , or of the c-statistic in a logistic regression.(Colinet et al., 2005; Hol-
man, Preen, Baynham, Finn, & Semmens, 2005) Figure 3 gives the average length of stay by level of
the Charlson Index. Note the considerable decrease for patients with an Index value of 13. Similarly,
there are drops at the Index value of 5 and 11. Again, an index of patient risk should be increasing as
the index increases.
It is clear that the highest level of the index has reduced risk and mortality compared to the previous
level. As there are only 12 patients in this group with an Index of 13, we can look at them in detail. All
12 have a diagnosis of HIV with a weight of 6. These results suggest that the weight for HIV should be
reconsidered; it is probably too high.(Zavascki & Fuchs, 2007) Figure 4 gives a comparison of the Index
to total charges. It again shows that the charges can decrease as the level of the Index increases.
Yet validation of an Index typically is declared by comparing the Index to patient outcomes, including
mortality, or by comparing it to another index.(Byles, D'Este, Parkinson, O'Connell, & Treloar, 2005;
Fried, Bernardini, & Piraino, 2003; Goldstein, Samsa, Matchar, & Horner, 2004; Groll, Heyland, Caeser,
& Wright, 2006; Lesens et al., 2003) While many factors can be correlated with mortality, the Charlson