Statistical Methods - Text Mining Techniques for Healthcare Provider Quality Determination - page 120

Table 3. Diagnosis codes used to predict mortality

Condition ICD9 Codes

Acute myocardial in-
farction

410, 412

Congestive heart failure 428

Peripheral vascular dis-
ease

441,4439,7854,V434

Cerebral vascular ac-
cident

430-438

Dementia 290

Pulmonary disease 490,491,492,493,494,495,496,500,501,502,503,504,505

Connective tissue dis-
order

7100,7101,7104,7140,7141,7142,7148,5171,725

Peptic ulcer 531,532,533,534

Liver disease 5712,5714,5715,5716

Diabetes 2500,2501,2502,2503,2507

Diabetes complications 2504,2505,2506

Paraplegia 342,3441

Renal disease 582,5830,5831,5832,5833,5835,5836,5837,5834,585,586,588

Cancer 14,15,16,17,18,170,171,172,174,175,176,179,190,191,193,

194,1950,1951,1952,1953,1954,1955,1958,200,201,202,203,
204,205,206,207,208

Metastatic cancer 196,197,198,1990,1991

Severe liver disease 5722,5723,5724,5728

HIV 042,043,044

The decision tree for a 25% sample (Figure 13) is considerably different from that in Figure 10 with a
50/50 split. Now, the procedure of Esophagogastroduodenoscopy gives the first leaf of the tree; in Figure
10, the first split was on age followed by charges and length of stay.Thus, a change in the sampling can
in and of itself be responsible for the outcomes predicted by the model.
Note that the trend shown in the 25% sample is even more exaggerated in the 10% sample. Figure 14
shows that the decision tree has changed yet again. It now includes the procedure of continuous positive
airway pressure and the diagnosis of congestive heart failure.

addItIon of WeIgHts for decIsIon makIng

In most medical studies, a false negative is more costly to the patient compared to a false positive. This
occurs because a false positive generally leads to more invasive tests; however, a false negative means
that a potentially life-threatening illness will go undiagnosed, and hence, untreated. Therefore, we can
weight a false negative at higher cost, and then change the definition of a “best” model to one that mini-
mizes costs. The problem is to determine which costs to use.
The best thing to do is to experiment with magnitudes of difference in cost between the false positive
and false negative to see what happens. At a 1:1 ratio, the best model is still based upon the misclas-