An unsupervised neural network approach to medical data mining techniques

This paper discusses a medical data mining technique introduced previously (DeClaris et al. (1996)) which combines unsupervised neural networks with data visualization. In this paper we present the application of this approach on a set of conventional pathology data. Inherent difficulties in the utilization of such data were overcome by utilizing three data subspaces identified as drugs, topography and morphology. A series of unsupervised neural networks were designed and actual data were used to evaluate their performance in identifying natural clusters of patient populations. Included is a method to examine and validate the underlying reasons for clustering. Preliminary examinations of identified clusters by qualified pathologists have shown promising results, which supports the conclusion that the suggested methodology yields discoveries and medical interpretations that can eliminate or serve as alternatives to special purpose epidemiological studies.