Knowledge discovery processing and data mining in karyometry.

Objective: To present the rationale for applying different sequences of multivariate analysis algorithms to determine if and where, in the large and high-dimensional data space, events have led to change in karyometric features.

Study Design: Clinical materials and results from the analysis of 4 studies were used: the demonstration of chemopreventive efficacy of letrozole in a situation where only a small subset of cells is affected, the detection of a preneoplastic lesion in colorectal tissue, data processing to document clues that predict risk of recurrence of a bladder lesion and the use of metafeatures and second-order discriminant analysis in a study of efficacy of vitamin A in the chemoprevention of skin lesions.

Results: Evidence for chemopreventive efficacy was demonstrated in the first example only after processing identified the small subpopulation of affected nuclei in a study of breast epithelial cells. Detection of a preneoplastic development is linked to a progression curve connecting nuclei from normal tissue to nuclei from premalignant colorectal lesions. The prediction of risk of recurrence of papillary bladder lesions is possible by detecting changes in nuclei of a certain phenotype. Efficacy of vitamin A as a chemopreventive agent for skin cancer could be demonstrated with a dose-response curve after a second-order discriminant analysis was employed.

Conclusion: In none of these instances would the information of biologic interest have been revealed by a straightforward, single algorithmic analysis.