Text Mining and Patient Severity Clusters - Text Mining Techniques for Healthcare Provider Quality Determination - page 309

cancer patients are in cluster 7, related to anxiety disorders, which seems inherently reasonable. It is
even more prevalent for patients with metastatic cancer, and for patients with HIV. The reason some of
these diagnoses that define the Charlson Index are scattered across the diagnosis clusters is that different
diagnoses are used to define the clusters. Figures 16, 17, and 18 give the procedure clusters as they are
related to these same patient conditions.
Cluster 5 is so prominent and related to so many of the Charlson Index diagnosis, that it is possible
that it is into this cluster where some upcoding may occur. Patient with peptic ulcers do not appear to be
in this cluster very prominently, or surprisingly, for patients with liver disease. Cancer and metastatic
cancer have a significant proportion in cluster 9, which contains procedures related to chemotherapy.
comParIson to aPrdrg
In a similar manner, we can compare the text clusters to the APRDRG index. Table 9 shows the propor-
tion of patients in each diagnosis cluster by APRDRG mortality index.
As Table 9 shows, there are some similarities. There are also some benefits to the use of the diagnosis
clusters since most of the patients in APRDRG category 0 that cannot be classified, were classified in
the clusters. Figure 19 shows where similarities occur. Note that over 90% of the patients in cluster 1
cannot be classified using the APRDRG Index. Over 80% of the patients in cluster 3 are in APRDRG
category 2. However, the remaining clusters tend to cross all of the APRDRG levels.
A graph of the APRDRG severity index by diagnosis cluster gives a similar result. We again use
a predictive model to determine whether the clusters can predict the APRDRG mortality level. The
misclassification rate on the testing set is 25%, indicating that this model predicts slightly better than
predicting the Charlson Index levels. The memory based reasoning model is the best, but the decision
tree misclassifies at a rate of 28%. It indicates that the first split is on age, but the second split is based
on diagnosis codes followed by procedure codes (Figure 20).
Since we define both the APRDRG Index and the clusters as nominal, we can include the entire
APRDRG index, including the patients that cannot be classified with the value of 0.
comParIson to resource utIlIzatIon
We do a similar examination in Table 10 using the Disease Staging: Mortality Level.
The clusters have a range of 0 to 10% (in cluster 9) of the patients in the highest severity level. A total
of 92% of the values in diagnosis cluster 1 are in the mortality level of 0. Table 11 gives the comparison
by procedure levels.
Some of the clusters are scattered across the mortality levels, while others such as cluster 6 are con-
centrated within one mortality level. We again want to see how well the clusters can predict the resource
level. We attempt to predict the disease staging: mortality level. The regression model is optimal, but
the error rate is 58%.