The All Patient Refined Diagnosis Related Group - Text Mining Techniques for Healthcare Provider Quality Determination - page 217

Charlson value of 13 corresponds to the APRDRG value of 4. However, there is a small percentage
of Charlson value zero patients that have an APRDRG index value of 4; this shows that there can be
considerable disagreement in defining the most severe patients.
We next look at the patient diagnoses that are used to define the Charlson Index (Figures 15, 16, and
17). For acute myocardial infarction (AMI) and congestive heart failure (CHF), there is a considerable
percentage of patients in the most severe category. Almost half of the CHF cases are in severity class 4.
No CHF patients are in class 1. Given the APRDRG distribution for CHF, it is very similar to that for
severe liver disease. It should probably have a higher weight in the Charlson Index.
Peripheral vascular disease (PVD) has a much lower proportion contained within APRDRG class
#4, as does a cerebral vascular accident (CVA) in spite of the fact that they have identical weights in
the Charlson Index, indicating a similar impact on patient outcomes. Almost half are in class 2 with a
small proportion in Class 1. The next two diagnoses of connective tissue disorder and peptic ulcer look
very similar. It indicates that these four probably should have similar weights in the Charlson Index or
that the Charlson Index should be re-evaluated. While diabetes also looks similar, liver disease has a
much higher proportion of patients in APRDRG index 4 with half of the patients in APRDRG index 3,
and no patients in index 1.
In fact, liver disease results look very similar to those in the pie charts for diabetic complications
and paraplegia. It should probably be given the same weight of 2 in the Charlson Index instead of the
current weight of 1. At the same time, there is little difference in the APRDRG index for liver disease
and severe liver disease. Perhaps they should be combined and receive a Charlson Index weight of 3
rather than 2. Metastatic cancer and HIV have similar distributions of the APRDRG index. However,
both appear to be less severe than liver disease. The surprise is that in the distribution for cancer, the
risk appears similar to that of AMI (acute myocardial infarction) and diabetes.
We next want to see if the APRDRG index, either alone or in combination with the Charlson Index,
can better predict patient outcomes compared to just the Charlson Index. We also want to see if the
Charlson Index can predict the APRDRG levels for this specific condition. In addition to the Charlson
index, we use patient demographic information to determine whether we can predict the APRDRG
mortality index. We use a predictive model on a 10% sample of the NIS dataset.
Figure 18 shows the variable roles for the predictive model. Since the outcome variable is ordinal
rather than binary, not all of the predictive models are compatible. Figure 19 shows the predictive model
used. Figure 20 gives the misclassification rate, indicating that regression is the optimal model choice.
Figure 21 gives the lift function.
In this model, we use age, race, and gender as well as income quartile. We also include length of
stay and total charges. ASOURCE, ATYPE, and AWEEKEND indicate the type of hospital admission.
We also use the Charlson Index to predict the APRDRG mortality index. ASOURCE indicates whether
the patient was admitted from the emergency department. ATYPE indicates whether the admission was
elective or not. AWEEKEND indicates whether the admission occurred on the weekend. Generally,
elective admissions do not occur on the weekend.
Note that there is little difference in the misclassification rates for 4 of the 5 models used. It indicates
that the model is approximately 70% accurate when classifying patients into the APRDRG mortality
index. However, the lift function indicates that it is primarily the top decile that can be predicted ac-
curately; starting in the second decile, the lift function is equal to one, indicating that the model does
no better than random chance. We also want to know the direction of the misclassification. Therefore,
except for these first two deciles, there is no real relationship in the assignment of severity between the