Information Technology Reference
In-Depth Information
already associated with the principal diagnosis. For example, a secondary diagnosis of urinary retention
is eliminated from a primary diagnosis of prostate hypertrophy. The next step is to assign each secondary
diagnosis to its standard severity of illness level. For example, uncomplicated diabetes is considered
minor while diabetes with renal manifestations is considered moderate. Diabetes with ketoacidosis is
major and diabetes with a coma is extreme. Next is to modify the severity of illness level based upon
the patient's age. It may also be modified based upon the principal or secondary diagnosis. For example,
renal failure increases the level of severity for patients with diabetes compared to patients with diabetes
but no listed complications. Physician panels were used to decide upon the level of severity for each of
the primary and secondary diagnoses.
While the predictive ability is small for patient outcomes and the APRDRG should not be used to
compare the quality of providers, the APRDRG can be used to identify providers who routinely treat
more severe patients.(Horn, Bulkley et al. 1985; Shen 2003) It can also be used to identify resource
utilization by specific groups of patients.(Ciccone, Lorenzoni et al. 1999) Since the APRDRG is very
sensitive to the assumption of the uniformity of data entry, it can be easily “gamed” by providers who
want to upcode to demonstrate that they have more severe patients. Studies have been conducted to
examine the relationship of the APRDRG to patient outcomes.(Shen 2003) It was discovered that the
APRDRG is highly sensitive to the quality of the data, and that model results can change considerably
with changes in the quality of data collection independent of the quality of patient outcomes.(Ciccone,
Bertero et al. 1999) Unfortunately, the relationship is often statistically significant because of the large
number of observations, but has very low predictive capabilities.(Rosen, Loveland et al. 2001; Naessens
and Huschka 2004)
comParIson of aPrdrg Index to PatIent outcomes
We would expect patients with a higher APRDRG index to have worse outcomes. Therefore, we examine
the three basic outcomes of mortality, length of stay, and total charges. For length of stay and charges, we
use kernel density estimation. Table 1 gives the relationship of mortality to the APRDRG severity index;
Table 2 gives it for the APRDRG mortality index. Figures 1-4 give the relationship of the APRDRG
index to the continuous outcomes. Table 3 compares the severity index to the mortality index.
Tables 1 and 2 do show an increase in mortality as the index increases, to a proportion of 32% in
mortality level four. The level of mortality for the severity index is lower compared to that of the mortality
index. A higher proportion of patients are in levels 3 and 4 for mortality compared to severity. Also, ap-
proximately 4000 or 0.05% of the patients cannot be classified into any level of the APRDRG index.
For those patients who can be classified, most of the patients are classified as mild followed by moder-
ate. Only a very small number are classified as extreme. As suggested in Chapter 3, one of these values
has to act as a threshold in a logistic regression. Very likely, the threshold that will be used is level 4.
However, these indices will predict mortality at a much higher rate compared to actual mortality since
the overall mortality is under 4%, but there are 32% of the patients in this category.
A code of zero indicates that the patient cannot be classified. Otherwise, the actual mortality increases
as the APRDRG index level increases as well. Therefore, before we examine a logistic regression, we
should eliminate all patients with an index level of zero. We filter these patients out of the model. Then
we will get an equation such that
Search WWH ::
Custom Search