The All Patient Refined Diagnosis Related Group - Text Mining Techniques for Healthcare Provider Quality Determination - page 213

proc kde data=nis.sample_addseverity;
univar los/gridl=0 gridu=10 out=nis.kdeaprmortalitylos1 method=os bwm=1.5;
univar totchg/gridl=0 gridu=30000 out=nis.kdeaprmortalitychg1 method=os bwm=1.5;
by aprdrg_risk_mortality;
where aprdrg_risk_mortality>0 and aprdrg_risk_mortality<4;
proc kde data=nis.sample_addseverity;
univar los/gridl=0 gridu=10 out=nis.kdeaprmortalitylos2 method=os bwm=.2;
univar totchg/gridl=0 gridu=30000 out=nis.kdeaprmortalitychg2 method=os bwm=.2;
by aprdrg_risk_mortality;
where aprdrg_risk_mortality=0 or aprdrg_risk_mortality=4;
run;
Figure 11 gives the length of stay by the APRDRG severity index. It shows that there are differences
in length of stay by index level. Figure 12 gives a similar graph for total charges.
As the severity index increases, the likelihood of a longer length of stay increases, with a crossover
point at 5 days. Patients in class 4 have considerable variability, so that there is no discernable pattern in
the outcomes. That occurs because there are so few patients in that class. Similarly, class 0 is highly vari-
able, which is reasonable since the patients in the category cannot be classified into a severity level.
The graphs in Figure 12 represent a similar result with the higher class having a higher probability
of higher total charges. Again, class 4 has considerable variability in its distribution as does class 0.
Figures 13 and 14 examine the Mortality Index. The distributions are relatively similar compared to
those for total charges.
Figure 11. APRDRG severity index and length of stay
Figure 12. Severity index and total charges