Background: Cluster analysis is particularly effective in detecting homogeneous subgroups among large series of observations. We applied this relatively uncommon approach to the study of prognosis in 137 patients affected by acute myeloid leukemia (AML).
Methods and results: Employing simple presentation parameters (age, WBC, splenomegaly, hepatomegaly) we used cluster analysis to define 3 groups with different overall survival (p = 0.0019). This classification was obtained following a rescaling of the variables and principal component analysis. Validation was performed through random definition of a control group. With the same variables, univariate analysis demonstrated age was the only prognostic factor, while Cox's model was not significant.
Conclusions: In our series cluster analysis allowed a better definition of prognosis than Cox's analysis. Since the 3 groups are well identifiable, each patient can be rapidly classified and his allocation confirmed by discriminant functions. For cluster 2 we were able to project a possible myelodysplastic evolution, while cluster 3 was more frequently associated with a monocytic blastic component. We think that cluster analysis deserves consideration as an alternative statistical approach in the analysis of large series of data; its usefulness lies in its power to define homogeneous prognostic or biologic subgroups and to elaborate further hypotheses for new studies.