Use of Risk Adjustment Models for Provider Reimbursements - Text Mining Techniques for Healthcare Provider Quality Determination - page 372

Figure 5. Length of stay by number of clusters
Figure 6. Total charges by number of clusters
Most of the codes are relatively similar in the two groups, indicating that the patients may not be that
far apart in terms of severity. After examining the most common codes, we use text mining to cluster
the codes separately for each group to see if they generate similar clusters. Diabetes is higher on the
frequency list for patients with 5 or more diagnosis codes compared to patients with 4 or fewer as is the
condition of COPD. These suggest partially that patients with 5 or more codes are sicker compared to
those with fewer than 5 codes. However, as many appear to have similar conditions, it seems that quite
a few patients were shifted into higher categories by adding codes.