Introduction to Ranking Models - Text Mining Techniques for Healthcare Provider Quality Determination - page 24

demonstrate how the information is entered (Figure 11).
Decision one is to predict mortality; decision zero predicts non-mortality. Currently, the cost of error
is the same, whether that error is a false positive or a false negative. When the weights are symmetric
as they are in Figure 11, the problem defaults to one of misclassification.
If we want to predict those at highest risk for mortality so that interventions can be used to prevent
it, it would be better to have more false positives, which will increase the cost of treatment, rather than
false negatives, which will result in greater unexpected mortality. In this case, we can change the values
to reflect the cost that we want to give. A positive value indicates a profit; a negative value indicates a
cost. Therefore, for medical treatment, we want to have a negative value to indicate the cost. Then there
is the option of whether to maximize or to minimize. We maximize profit when the weights are positive;
we minimize cost when the weights are negative.
The NIS and the MEPS datasets have two different types of listings of diagnosis and procedure
codes. The NIS data have 15 columns of both procedures and diagnoses for each patient. They can eas-
ily be combined using the CATX function in SAS. The MEPS data include a patient condition file with
multiple conditions per patient and one condition per observation. Combining these conditions into one
observation per patient requires a bit more coding. While the coding will be discussed in more detail in
later chapters, it will be examined briefly here.
nIs data
The following code adds two new columns to the dataset:
Data NIS.combinecodes;
Set nis.datafor2005;
Diagnoses=catx(' ',dx1,dx2,dx3,dx4,dx5,dx6,dx7,dx8,dx9,dx10,dx11,
dx12,dx13,dx14,dx15) ;
procedures=catx(' 'pr1,pr2,pr3,pr4,pr5,pr6,pr7,pr8,pr9,pr10,pr11,pr12,
pr13,pr14,pr15) ;
run ;
Table 6 shows the results of the above code for the diagnoses. The result is to create a text string
consisting of all the codes for an individual patient.
In order to make sense of Table 4, we need to translate these ICD9 codes. We use the online informa-
tion at http://icd9cm.chrisendres.com. The codes are given in Table 7.
The first four patients in the list suffer from pneumonia. However, patient #1 also suffers from asthma
and obesity. Patient #2 has an additional upper respiratory infection while patient #3 also suffers from
asthma and from high cholesterol. Patient #5 suffers from severe heart problems, including a previous
heart attack (myocardial infarction). Patient #6 has Moyamoya disease, a rare, progressive cerebrovascular
disorder. Along with this disease, the patient had a blackout and hemorrhage. In addition to pneumonia,
patient #7 has problems related to kidney disease and anemia. Patient #8 has diabetes and cellulitis in
the leg due to staphylococcus. In addition, this patient has anemia and pneumonia. Number 9 has severe
hypertension, edema, and diabetes. Along with pneumonia, patient #10 is dehydrated with asthma and
pleurisy, having had a previous heart attack. Patient #11 has severe heart problems as well as a sexually
transmitted disease, and was previously a liver donor. Patient #13 has kidney failure, heart failure, and is