Use of Risk Adjustment Models for Provider Reimbursements - Text Mining Techniques for Healthcare Provider Quality Determination - page 376

Table 12. continued

Cluster
Codes for
4 or Fewer
Codes

Translation of Codes Cluster

Codes for
5 or More
Codes

Translation of Codes

6823, 6822,
04111, 6827,
99859, 6825

Other cellulitis and abscess (Upper arm and
forearm), Other cellulitis and abscess (Trunk),
Staphylococcus aureus, Other cellulitis and
abscess (Foot, except toes), Other postop-
erative infection, Other cellulitis and abscess
(Buttock)

V090, 6826,
2181, 6827,
04111, 4240

Infection with microorganisms resistant to penicillins,
Other cellulitis and abscess (leg, except foot and ankle),
Intramural leiomyoma of uterus, Other cellulitis and
abscess (foot,except toes), Staphylococcus aureus, Mitral
valve disorders

2761, 2768,
5990, 5849,
42731,
49121

Hyposmolality and/or hyponatremia, Hypo-
potassemia, Urinary tract infection, site not
specified,Acute renal failure, unspecified,Atrial
fibrillation, Obstructive chronic bronchitis

V4385,
v5481,
43820,
v5789, 7812,
v4364

Organ or tissue replaced by other means,Aftercare following
joint replacement, Hemiplegia affecting unspecified side,
Other specified rehabilitation procedure, Abnormality of
gait, Organ or tissue replaced by other means (hip)

For the first group, the cellulitis is likely to be from staph, but not resistant staff; in the second group,
the infection is from resistant staph. In addition, the first group is more likely to have the infection in
the arm, trunk, or buttocks whereas the second group has the infection in the foot and leg. The locations
indicate that the first group has more community-acquired infection in addition to hospital acquired post-
operative infection while the second group is more likely to have hospital-acquired infection. Moreover,
the second group has related heart problems.
Table 12 shows that patients with 5 or more codes are quite simllar to patients with 4 or fewer codes,
with just slight differences in the actual codes themselves. More drill down to examine the nature of
these codes in more detail will find the nature of the gaming.

examPle of cardIovascular surgery

To further examine the aspect of “gaming”, we look at the number of diagnoses by hospital when restricted
to the cardiovascular procedures in 36.1. Table 13 shows the proportion of patients with the number of
diagnosis codes for each procedure. Note that the proportion skews towards a higher number of codes.
Procedure 3615 has the highest proportion of patients with 15 diagnosis codes.
Table 14 gives the relationship of number of codes to hospital. Hospital #1 has almost 60% of the
patients with all 15 diagnosis codes while hospital #7 has none; the most these patients have is 14 at 65%.
As we have already shown, hospitals that “game” the system tend to identify more codes per patient.
We use a predictive model to see if the number of codes increases the proportion of predicted mor-
tality. Table 15 gives the proportion of patients who died by the number of diagnoses. Note that all of
these patients have at least 12 different diagnoses. Since the diagnosis codes are assigned at discharge,
it is possible that the number of diagnoses was increased after death.
Figure 7 gives the results of the predictive model, which includes hospital, procedure, and number
of diagnoses without any patient demographics; the memory based reasoning node is the best model.
Table 16 gives the quality ranking based upon the number of diagnosis codes. Table 17 compares the