Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 15

Table 5. continued

Data Element Variable Name Description

Disease Staging (Medstat) DS_Mrt_Scale Mortality scale

Disease Staging (Medstat) DS_RD_Level Resource demand level

Disease Staging (Medstat) DS_RD_Scale Resource demand scale

Linkage Variables HOSPID Hospital identification number

Linkage Variables KEY Patient identification number

We will be discussing these data elements in later chapters.

claims data

Claims data is usually messier compared to the MEPS and to the NIS data. Unlike these two datasets
provided by the government, claims data are drawn from a variety of sources, including hospitals, physi-
cian offices, laboratories, radiology centers, and pharmacies. The different sources use different coding
systems, with hospitals relying primarily upon ICD9 codes and physician offices using CPT codes.
Pharmacies do not generally provide diagnosis information with their medications.
A severity index using claims data must be able to handle the different coding schemes simultaneously.
We will demonstrate how this can be done in Chapter 9. In addition, claims related to just one treatment
episode, including follow up visits needs to be isolated within the data in order to investigate patient
outcomes. Because claims are for specific services, episodes of treatment are not clear in the data.

data mInIng and HealtHcare dataBases

Clinical trials have always represented the gold standard in healthcare. However, there are limitations
that include cost, time, question, randomization, and sample size. Clinical trials are expensive to conduct;
therefore, not every question concerning medical decisions, new drug treatments, and surgical proce-
dures can be answered with clinical trials. Also because of the expense, clinical trials tend to examine
short term effects only, often relying upon surrogate rather than actual endpoints, and then making the
assumption that the surrogate is exactly related to the actual patient outcomes. Clinical trials also focus
on one treatment at a time, usually excluding subjects who are taking combinations of medications. Also
because of the expense, clinical trials do not enroll enough subjects to detect rare occurrences of adverse
events. Moreover, clinical trials must rely upon volunteers. If the treatment is for a life-threatening ill-
ness that cannot be treated with currently available methods, the majority of patients will undoubtedly
volunteer. However, in other treatments, there can still be a self-selection bias in the volunteer sample.
Therefore, we cannot always be certain that a randomized, controlled clinical trial is unbiased.
When the gold standard is not available, we must rely upon information that we do have for analysis
purposes. Clinical and administrative databases that are collected routinely have considerable value in
terms of patient treatments and outcomes. However, these data are observational, and observational data
can always introduce the possibility of confounding factors. Potential confounders need to be examined