Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 10

Both methods are limited because panels may reach consensus without examining the validity of
the choices. If one provider has very severe patients with comorbidities that are not included in the list,
these patients will appear to have a very low severity level.
In an interesting study concerning the use of high deductible/medical savings plans, Humana used
the utilization of healthcare services as identified through claims data to segment its employees into risk
clusters to compare the introduction of a new type of insurance. It was found that those who utilized
services less often were more likely to switch to the new plan.(Tollen, Ross, & Poor, 2004) Another study
used the incidence of a ruptured appendix in children as a marker for the quality of care. However, there
was no discussion of how to validate the use of just one diagnosis factor to rank the provider quality.
(Gadomski & Jenkins, 2001)
Given patients with similar conditions, we should be able to rank the quality of hospitals and other
medical providers by the outcomes of their patients. However, there is considerable variability in patient
conditions, so much so that it is rare to have similar patients at different hospitals. Numerous attempts
have been made to develop a method to equate patients at specific levels of risk so that outcomes can
be compared. In many cases, these measures are used to define those providers of highest quality, and
to rank the providers in order of quality. They are also used to compare different providers directly to
see if one is better than another.(Reilly, Chin, Berkowitz, Weedon, & Avitable, 2003)
Other measures of patient severity are used to determine those patients at highest risk, and those
who will be the most costly to treat, so that reimbursements to hospitals and providers can be adjusted
by risk levels. (Macario, Vitez, Dunn, McDonald, & Brown, 1997) However, different measures of pa-
tient severity can lead to different rankings of provider quality, making these rankings of questionable
validity. We should expect that rankings are invariant of the measure used to define the rankings. In
the absence of such consistency, the validity of ranking may be restricted to the top and bottom of the
ranking list. (Daley, Ash, & Iezzoni, 2003) In the absence of such consistency, we need to understand
the limitations of defining risk-adjusted quality, and to be careful when assigning reimbursements based
on the outcomes of these models.
Still other measures were developed to predict patient outcomes, and to develop standard measures
for specific patient populations.(Monami et al., 2007) This seems to be reasonable for some procedures,
but not for others. Can we define one risk adjustment method that works for heart patients and elderly
patients with dementia, but also for labor and delivery patients? Should risk be age, race, and gender
specific? This is probably not reasonable, which is why a risk adjustment model is defined for a specific
patient procedure.
Some measures are defined as composites of other measures to see if the multiple measures predict
better than just using one measure.(Azarisman et al., 2007; Inouye et al., 1998; O'Connell & Lim, 2000)
Similarly, different risk adjustment measures are compared directly.(Rochon et al., 1996)
Generally, a set of patient diagnoses is assembled, and a different weight is assigned to each of the
diagnoses. Then, a linear combination of weights is defined for each patient, with the combination con-
sisting of only those defined diagnoses that apply to that patient. The question to ask is how the set of
diagnoses is defined. Different sets will result in different risk adjustments; often, the different methods
will contradict each other. In some cases, a patient defined as having low risk in one measure can be
defined as having high risk in another measure.(L. I. Iezzoni et al., 1995) Assigning weights presents
another problem. Should diabetes have higher or lower weight compared to kidney disease? If each di-
agnosis is given the same weight, then the risk adjustment method is essentially a count of the number
of diagnoses for each patient; again, the number is based upon a defined set of diagnoses.