In order to evaluate the discovered risk groups, the medical expert has tested the induced subgroup patterns on an independent set of 70 people (50 CHD patients and 20 non-CHD cases from the same hospital). The results for these patients, summarized in Table 4, show that the patterns are successful in detecting CHD patients. About 90% of CHD patients were included into at least one of the five patterns. The detected sensitivity values (TPr) for patterns A1, B2, and C1 are significantly higher than the values computed on the set of patients used for subgroup discovery. For the other two patterns the values do not differ significantly. Note that the accuracy values are relatively high, despite the relatively high false positive rate (FPr): a lower FPr could have been achieved by selecting lower values of the generalization parameter g, at a cost of detecting subgroups with lower coverage of positive cases.
|