Adaptive resonance theory with supervised learning and large database applications
Vassilis G. Kaburlasos · 1992
This work describes a novel database processing approach based on neural networks to examine patterns of health care services in relation to the outcomes of care. The neurocomputing paradigm selected is the fuzzy adaptive resonance theory (ART) which is systematized mathematically and enhanced through the addition of supervised learning as developed by Gail Carpenter and Stephen Grossberg. This enhanced ART neural network accepts analog input patterns which are grouped, by a biologically motivated process of forced self-organization, into categorical The input patterns and categorical codes are only instances in the set of all codes. This set is shown to be a fuzzy lattice with distance. Input patterns augmented by their complements are organized into categories as in a conventional ART neural network. However, the network reset is triggered when a maximum allowable category size is exceeded. By using two of these networks, coupled with an intermap layer, a new neural network is obtained (fuzzy ARTMAP) that can be trained to self organize cause-and-effect pairs into categories that correspond to the same effect by extracting common features. This new neural network is applied to medical databases in order to find patterns of care and to make predictions about outcome based on a case-by-case learning experience.