085-2007: Text Mining and PROC KDE to Rank Nominal Data

By definition, nominal data cannot be ranked. However, there are circumstances where it is essential to rank nominal data. Examples of such ranking include ranking hospitals and colleges, defining the “most livable cities”, and conference paper submissions. In this project, we consider ranking patient severity. The purpose is to determine how patient severity can be used to rank the quality of hospital performance. There are thousands of patient diagnoses and co-morbidities that make such a ranking very difficult. Generally, nominal variables have been ranked by using quantitative outcome variables. Currently, hospital quality measures used stepwise logistic regression to reduce the number of patient diagnoses considered to define a measure of patient severity. More recently, a weight-of-evidence method has been developed for predictive modeling such that nominal data are compressed and ranked using a target variable. However, there are now methods available that allow for ranking nominal data that do not require outcome variables; instead, outcome variables can be used to validate the ranking. Ranking can be done using SAS Text Miner to compress nominal data fields containing information on patient diagnoses, combined with PROC KDE to define and validate the patient severity ranking. It will be demonstrated that SAS Text Miner can define an implied ranking of nominal fields that is identified through the application of PROC KDE. Once the patient severity rank has been defined, it will be used to examine patient outcomes, and physician variability in patient outcomes.