Text Mining and Patient Severity Clusters - Text Mining Techniques for Healthcare Provider Quality Determination - page 334

Table 23. Scored diagnosis clusters for 2004 data

Cluster Number Number of MRSA Patients Predicted into

Cluster

Cluster Number Number of MRSA Patients Predicted into

Cluster

1 98 17 365

2 623 18 64

3 66 19 100

4 191 20 320

5 71 21 94

6 48496 22 134

7 447 23 72

8 1971 24 700

9 86 25 78

10 225 26 469

11 260 27 546

12 145 28 477

13 210 29 2680

14 105 30 27

15 55 31 181

16 316 32 451

Diabetes with other specified manifestations (of unspecified type), Other bone involvement in diseases
classified elsewhere. While these patients are not at the highest level of MRSA, they are at a very high
level. It does suggest that at the least, infection control procedures should be used on this class of patients
to reduce the occurrence of MRSA. Another large number of infections are predicted into cluster 29,
Excessive or frequent menstruation, Other specified rehabilitation procedure, Acute posthemorrhagic
anemia, Organ or tissue replaced by other means, knee joint, Iron deficiency anemias, secondary to blood
loss, Diverticulosis of colon (without mention of hemorrhage). These patients, however, have a much
lower occurrence of MRSA and prophylactic treatment is probably unnecessary.
We next use predictive modeling. First, we need to define a profit/loss matrix to determine the differ-
ence between a positive prediction of MRSA, and a negative prediction. Then, we need to partition the
sampled dataset into training, validation, and testing subsets. We also use several models to determine
the model of best fit.(P. B. Cerrito, 2007) Since we have over-sampled the occurrence of MRSA, we also
need to set the prior probabilities of occurrence and non-occurrence to those that exist in the population.
In this first example, the procedure and diagnosis clusters are defined as interval variables.
The use of a profit/loss matrix will not change the model results; it will change the choice of which
model is optimal. The default for “best” fit is the misclassification rate; this can be modified to finding
the model with the highest profit (or minimal loss). Therefore, by changing the profit/loss matrix, we
can perform a sensitivity analysis to determine when the optimal model choice will change. We use the
available patient demographics as well as the categories of patient diagnoses and procedures. Using
just the misclassification, the optimal model is a regression with a 50% misclassification rate. Since
we used a 50/50 split in the data, 50% misclassification is a poor fit. If the cost of treatment is 10 times
the cost for prophylactic treatment, the best model is a neural network with an overall average cost less