Artificial neural networks: from black-box to grey-box modelling

The human decision-making process is simultaneously (and sometimes contradictory) logical, associative, fuzzy and probabilistic. To model this process faithfully, computer-aided medical diagnostic systems should possess all of these features. Therein lies one of the problems which hinders successful development of computer-based diagnostic aids: none of the software tools available has complete `coverage´ of the required features. One solution is to use more than one modelling paradigm to increase coverage. This paper describes a hybrid artificial neural network-causal probabilistic network modelling paradigm, developed to aid in the management of septicaemia. An emergent property is a method to visualise artificial neural networks in terms of the probability distributions of the causal probabilistic networks