Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 139

Table 10. Patients with pneumonia by hospital

Table of DSHOSPID by pneumonia

Hospital Code pneumonia Total

Frequency
Row Pct
Col Pct

0 1

1 2638

91.66
9.14

240
8.34
11.23

2878

2 1382

91.77
4.79

124
8.23
5.80

1506

3 830

91.51
2.87

77
8.49
3.60

907

4 7416

94.75
25.68

411
5.25
19.22

7827

5 2273

93.93
7.87

147
6.07
6.88

2420

6 4802

88.79
16.63

606
11.21
28.34

5408

7 1416

94.09
4.90

89
5.91
4.16

1505

8 932

99.36
3.23

6
0.64
0.28

938

9 5149

94.56
17.83

296
5.44
13.84

5445

10 2038

93.49
7.06

142
6.51
6.64

2180

Total 28876 2138 31014

By this process, hospitals #8 and #10 have the smallest differential between the actual and predicted
values. Therefore, they would be ranked the lowest even though they both have very low mortality
values. In contrast, hospital #6 with the highest actual mortality would rank the highest because the
difference between the actual and predicted mortality rates is the greatest. However, a hospital with 0
actual mortality has very little room for an increase in the predicted mortality; a hospital with a higher
actual mortality is more likely to “game” the system by increasing the predicted mortality.
Our second example is restricted to the treatment of patients with a primary diagnosis of COPD. We
use a predictive model that is similar to that in the previous section, but now we add a hospital identi-
fier. The results are given in Figure 35 with the ROC curve in Figure 36. Note that the minimum error
rate is still 32%.