Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 11

Once a method of risk adjustment is defined, a logistic regression is used to see if patient mortality
outcomes can be predicted.(Cher & Lenert, 1997) Linear regression is used to predict patient costs, and
also length of stay. (Shwartz & Ash, 2003) That is the reason that only a small number of patient diag-
noses can be used to define the score; logistic regression cannot use all of the thousands of codes that
are assigned to patients. However, while p-values are statistically significant, the correlation coefficient,
or r 2 value, tends to be low in a linear regression, suggesting that most of the variability in the outcome
is not predicted by the severity of patient condition. (Rochon et al., 1996) Logistic regression will have
a high accuracy level, reflecting only the fact that mortality is a rare occurrence. This rare occurrence
is almost never taken into consideration when developing the model, when some model adjustments
are required.
One of the more public risk adjustment measures is available at healthgrades.com. There, you can
find rankings of hospitals all over the United States. For a fee, you can also get a report on any physi-
cian. The developers at healthgrades.com discuss their methodology at www.healthgrades.com/media/
DMS/pdf/HospitalQualityGuideMethodology20072008.pdf so that we can examine in detail some of
the problems as well as the benefits of using this methodology. As discuss in their methodology, they
do use a logistic regression, and the document provides the patient diagnoses used to examine patient
outcomes. In some cases, healthgrades.com examines specific procedures such as cardiac bypass, and
then creates a risk adjustment for the remaining general patient population. For most of their ratings,
their measures examine mortality. If the procedure has virtually no mortality, then it ranks based upon
the occurrence of complications.
Another consideration in defining provider quality is the proportion of medical errors and complica-
tions resulting from an inpatient stay. If a hospital has a high rate of nosocomial infection, then it should
probably not rank as high in terms of quality with a hospital with a much lower rate of nosocomial
infection. Therefore, in addition to risk adjusted severity models, additional measures should also be
computed and compared by provider.
datasets used
Throughout this text, we will be relying on two different databases. The first is the Medical Expendi-
ture Panel Survey (MEPS). The second is the National Inpatient Sample (NIS). The MEPS is available
through a download from http://www.meps.ahrq.gov/mepsweb/data_stats/download_data_files.jsp. It
is very complete data for a cohort of patients. It is also available for ten years and contains informa-
tion on inpatient and outpatient treatment as well as physician visits, medications, and detailed patient
demographics. The NIS is available for a small fee from http://www.hcup-us.ahrq.gov/nisoverview.jsp.
It contains all inpatient visits from a stratified sample of hospitals from a total of 37 different states. In
addition, we will discuss some proprietary data as needed. In particular, we will discuss how Claims
data can be used to determine patient severity. The two databases that are publicly available represent
the different types of information that can be found in claims data, or in the electronic medical record.
mePs data
In the case of MEPS data, the datasets are available in SAS format, or in SPSS format. SAS and SPSS
coding are provided to translate the datasets into usable files. In addition, data dictionaries are readily