Adaptive Resonance Theory: A Foundation for "Apprentice" Systems in Clinical Decision Support?

The idea of an "apprentice" system in contrast to an expert system, is introduced, as one which continues, perpetually, to refine its knowledge-base. Neural networks appear to offer the necessary learning ability for this task, and the Adaptive Resonance Theory family is particularly suited to on-line (casual) learning. The ability of these networks accurately to represent decision problems and to disclose their acquired knowledge is discussed, and their practical application is assessed. Two problems of medical decision making are considered using the approach. The first is the early diagnosis of myocardial infarction from clinical and electrocardiographic data gathered at presentation. The second is the cytopathological diagnosis of breast lesions from fine needle aspirate samples. In both cases good performance is obtained along with sets of "if-then" rules which are in accordance with medical opinion. In the first case, examples of on-line learning are given and the system is seen to be behaving as expected, with performance improving with increasing sample size.