The rapid classification of brain conditions using neural networks

It has been shown (Jervis et al. 1994) that it is possible to differentiate between patients suffering from Huntington´s disease, Parkinson´s Disease, or schizophrenia and normal healthy individuals on the basis of an evoked response in their electroencephalograms known as the contingent negative variation (CNV). To some extent it is also possible to differentiate between these conditions. This has been done several ways, but most recently by pre-processing the recorded CNV data, extracting certain time domain features and using the normalised features as the input vectors to artificial neural networks (ANNs) (Jervis et al. 1994). This type of medical data is only accumulated slowly and the data sets are likely to be small. Also clinicians are more likely to accept such new methods based upon computer use if they have control over the data entered into the computer. It is desirable to use an ANN which could easily be trained by the clinician, which is insensitive to the relative numbers of training vectors of each class, which is capable of incremental training in between testing sessions, and relies upon few unknown coefficients which have to be chosen by the user. The simplified fuzzy ARTMAP (SFAM) (Kasuba, 1993) is such an ANN. Derived from the fuzzy ARTMAP, there is only one significant adjustable parameter, known as the vigilance, which determines the classification resolution. The SFAM is purely a classifier. It is desirable to have some knowledge about the reliability of the classification and for this reason two networks called the probabilistic simplified fuzzy ARTMAP (PSFAM) and the integrated PSFAM were developed.