Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 2

Because there are so many possible patient conditions, the results of risk adjustment are more likely
to reflect the definition of the model rather than to reflect actual comparisons of patient severity. No
matter what patient conditions are included in the model, there are many more excluded that could be
just as crucial when considering patient risk. Moreover, the few that are included will cover only a small
percentage of patients.
In addition, risk adjustment models only consider patient condition and not patient compliance with
treatment.(Rosen, Reid, Broemeling, & Rakovski, 2003) This paper suggests that health status is dependent
upon health behaviors and psychosocial factors as well as the social environment and socioeconomic
status of the patients themselves. Therefore, a physician with more lower-income and minority patients
will have health outcomes that are not as strong as a physician with mostly affluent patients. However,
that brings up another issue. Just how should health behaviors be identified and ranked? In other words,
risk is an extremely complex issue that has multiple dimensions, and all dimensions contribute to risk.
Without looking at all of these factors and dimensions, risk adjustment models will continue to be
questionable.
Moreover, any model that defines risk should be subject to strict scrutiny to determine its validity.
Otherwise, it is possible to decide that heart patients are sicker than cancer patients, and both are sicker
than dialysis patients. The degree of “sickness” is usually defined in a model by assigning weights to
each condition. The greater the weight, the greater the sickness. However, this proposed model of heart,
career, and dialysis patients uses three conditions only. Would it be a better model to include pneumonia
and asthma? At some point, every model includes some conditions but excludes others. It is possible
that excluded conditions can be more severe than included conditions.
In this text, we will discuss some common methods for defining patient severity and compare results
using different models. In addition, we will propose a technique that uses all patient diagnoses and pro-
cedures to define a patient index. Since there are many different methods of risk adjustment, the different
methods can give very different results.(L. Iezzoni, Shwartz, Ash, & Mackieman, 1994) However, if the
results can be so different, how can any risk adjustment model be validated? Indeed, can it be validated
when the results can be so different?
It is not enough to simply use a statistical measure to claim validation because of a number of prob-
lems, including over-fitting by including too many diagnoses. Another problem is caused by using too
few diagnoses; unfortunately, both problems commonly occur.(Singh et al., 2003) A previously published
edited text on risk adjustment discusses at length several indices with different choices of patient diag-
noses. The chapter written by Lisa Iezzoni states that patients with comorbidities have higher risks of
death and complications as is logical, usually have higher rates of functional disability, and often require
additional diagnostic testing and treatment interventions. (Iezzoni, 2003) However, the text goes on to
say that using comprehensive criteria to specify diagnoses is not reasonable, and that it is not possible
to identify specific diagnoses usually because of a lack of knowledge or information. The number one
issue for any model of patient severity is how to handle all possible combinations of diagnoses.
Using publicly available data and coded information about patient conditions, quality can be de-
fined in many ways. However, it is primarily defined as the difference between predicted and actual
mortality, although the ratio of predicted and actual mortality can also be used. (Ash et al., 2003) Other
information might be given as well and used to modify the ranking. For example, the Texas Hospital
Checkup makes available the following information: (Anonymous-TGBH, 2008; Arca, Fusco, Barone,
& Perucci, 2006)