There is a requirement for a rapid classifier for clinical applications. Ideally it should be possible to build it quickly, to enlarge its structure incrementally, to be able to detect misclassifications, and to use it on-line. Such a classifier is described, together with an example application to differentiate between the Contingent Negative Variation evoked responses in tbe electroencephalogram of a number of subject groups. Known as the Probabilistic Simplified Fuzzy ARTMAP (PSFAM), it consists of a combined Simplified Fuzzy ARTMAP (SFAM) artificial neural network committee, and a Bayes' classifier, implemented with the aid of the Parzen windows technique. It has been shown to improve the classification accuracy compared with an SFAM committee alone. The SFAM committee is rapidly trained, requiting only two iterations of the training data, and may learn on-line. The posterior probability of belonging to a class obtained from the Bayes' classifier yields a measure of confidence in the classification In solving the demonstration medical diagnosis problem classification accuracies in the range of 86-96% were achieved, as wed as in many cases ideal or near ideal values of sensitivity, specificity, and false positive and negative rates. The classification performance of the PSFAM seemed to be limited by the nature of the data rather than by the method.
