Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 3

• Actual vs. expected mortality

Number of procedures done a year •

• Average length of stay

• Average total charges

• Average charge per day

Since providers are ranked by the difference between actual and predicted mortality, then, if a pro-
vider has zero actual mortality, and zero predicted mortality, then the difference is zero. For this reason,
a provider with zero mortality can actually rank lower in quality compared to providers that have high
levels of actual mortality. We will demonstrate this seeming contradiction in subsequent chapters.
However, if the magnitude of difference between actual and expected mortality is the measure of
quality, it is clear that providers can improve their quality ranking either by decreasing the actual mor-
tality, or by increasing the expected mortality. Increasing the expected mortality is easier, using what is
known as “upcoding”. (Yuill, 2008) That means that the provider can add diagnosis codes to describe
the patient condition that increases the level of identifiable risk. Although considered fraud, upcoding
remains widely practiced.(Lorence & Spink, 2004)
Accuracy in identifying risk is more important than ever since these risk adjustment models are used
to set health insurance premiums, and Medicare hospital payment rates as well as to measure provider
performance. (O'Leary, Keeler, Damberg, & Kerr, 1998). Because of the relationship of quality ranking
to reimbursements, there is substantial pressure to upcode patient conditions.
We also need to determine whether to focus on errors, or upon optimal care. We need to decide whether
to focus on inputs, meaning adherence to treatment guidelines, or upon outcomes. Because outcomes
are often difficult to examine, many measures tend to focus only upon inputs. Did the patient receive
necessary tests and medications given the defined diagnosis, or did the provider minimize conditions
that result in a spread of infection? However, ultimately, patient outcomes are more important than in-
puts, and we should concentrate upon outcomes. A focus on inputs assumes that the correct inputs will
result in the best outcomes. Not every decision or procedure is subject to guidelines; therefore, outcomes
depend on more than just an adherence to known guidelines.

Background In coded InformatIon

The datasets used to define patient severity indices rely upon patient conditions as identified through
coding systems and these codes must be integrated. There are a number of coding systems that are used
in administrative data. The first is the DRG, or diagnosis related group. These codes are used by Medi-
care and other insurance providers to provide reimbursements. Generally, a value is negotiated between
the insurer and the provider for a particular DRG. Generally, DRG grouper software uses “the principal
diagnosis, secondary diagnoses (as defined using ICD9 codes), surgical procedures, age, sex and dis-
charge status of the patients treated” to assign inpatient records to a specific DRG. (Anonymous-DRG,
2008b). Examples of DRG codes are listed below:(Anonymous-DRG, 2008a)

424 Operating room procedure with principal diagnoses of mental illness
425 Acute adjustment reaction & psychosocial dysfunction
426 Depressive neuroses