Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 138

Table 9. Patients with septicemia by hospital

Table of DSHOSPID by septicemia

Hospital Code septicemia Total

Frequency
Row Pct
Col Pct

0 1

1 2782

96.66
9.24

96
3.34
10.49

2878

2 1444

95.88
4.80

62
4.12
6.78

1506

3 892

98.35
2.96

15
1.65
1.64

907

4 7628

97.46
25.34

199
2.54
21.75

7827

5 2347

96.98
7.80

73
3.02
7.98

2420

6 5035

93.10
16.73

373
6.90
40.77

5408

7 1498

99.53
4.98

7
0.47
0.77

1505

8 938

100.00
3.12

0
0.00
0.00

938

9 5360

98.44
17.81

85
1.56
9.29

5445

10 2175

99.77
7.23

5
0.23
0.55

2180

Total 30099 915 31014

gives the rate for immune disorder. The trend is similar; hospital #8 has the lowest rate, hospital #6 has
the highest.
These three tables suggest that hospital #6 has a very good reason to have a higher mortality rate.
For this reason, we compare the expected mortality to the actual mortality. We use a predictive model
with hospital, septicemia, immune disorder, and pneumonia as the input variables and mortality as the
output variable. Figure 33 gives the results, indicating that Dmine regression gives the best fit.
The best misclassification rate is still almost 30%. We partition the data to define the model; we then
score the entire dataset so that we can examine the difference between the predicted and actual values.
Figure 34 shows the datasets generated by the score node in Enterprise Miner.
The dataset EMWS3.Score_Score contains the predicted values as well as the actual values. We can
use PROC FREQ in SAS to examine the relationship to hospital. Table 12 gives the actual and predicted
values by hospital.