How to Check Measures for Adequacy - Text Mining Techniques for Healthcare Provider Quality Determination - page 391

to relate payer reimbursements based upon rankings of quality, it is imperative that the risk adjustment
models used are validated so that the highest quality providers are the highest scoring in the models.
Because of these issues, the habit of using the difference between predicted and actual outcomes should
be abandoned in favor of a multi-dimensional approach.
When risk adjustment methods are used to reimburse providers based upon their ranking, the meth-
odology needs to be scrutinized very carefully. The potential for “gaming” should be investigated, and
the possibility of penalizing good quality providers is important, and should be examined carefully. Real
data should be used to consider the validity of the method without just considering model reliability.
Predictive modeling techniques are important when analyzing both validity and reliability, and should
be used along with the more traditional regression. When large datasets are used to define and examine
patient severity, the p-value is of little value and should not be used to measure the adequacy of the model.
Also, the rareness of the outcome measure should be considered, and compensated for using stratified
sampling. The measure of quality is a complex issue and will require some fairly complex modeling.
future trends
Because patient severity and hospital quality are both multi-dimensional, the most effective measures
of both will be multi-dimensional. Therefore, the use of ICD9 codes to define a patient severity should
be just one dimension to define quality. Patient compliance with treatment must also be included as a
dimension. Moreover, medical errors and nosocomial infections should be included as additional dimen-
sions. If these added dimensions are not considered, there will still be a problem of validity.
Data mining techniques allow us to drill down to explore details; such drill down is not possible
with standard statistical techniques. Statistical methods ultimately reflect group identity rather than to
examine individual identity. Therefore, data mining can find more and better detail in the data. These
techniques will certainly be used more often as more electronic data become available. While the transi-
tion to electronic records has been slow, this transition is continuing. The more electronic information
available, the more knowledge that will be extracted from those electronic records.
dIscussIon
The problem of defining a patient severity index is still not resolved. If it were, there would not be so many
different measures and different attempts to create severity models. Providers should take opportunities
to define their own measures of quality and to measure improvements in quality within themselves as
well as in comparison to other providers.
In this text, we have discussed these additional dimensions briefly in addition to examining the current
techniques available to define patient severity measures. In addition, we have demonstrated how data
mining techniques, particularly data visualization, predictive modeling, and text mining can be used to
define better measures of patient severity that do not require accepting false assumptions, especially the
assumption of the uniformity of data entry.