ARTIFICIAL NEURAL NETWORKS THAT USE SINGLE-PHOTON EMISSION TOMOGRAPHYTO IDENTIFY PATIENTS WITH PROBABLE ALZHEIMERS-DISEASE

Single-photon emission tomographic (SPET) images using technetium-99m labelled hexamethylpropylene amine oxime were obtained from 97 patient s diagnosed as having Alzheimer's disease, as well as from a compariso n group of 64 normal subjects. Multiple linear regression was used to predict subject type (Alzheimer's vs comparison) using scintillation c ounts from 14 different brain regions as predictors, These results wer e disappointing: the regression equation accounted for only 33.5% of t he variance between subjects. However, the same data were also used to train parallel distributed processing (PDP) networks of different siz es to classify subjects. In networks accounted for substantially more (up to 95%) of the variance in the data, and in many instances were ab le to distinguish perfectly between the two subjects, These results su ggest two conclusions. First, SPET images do provide sufficient inform ation to distinguish patients with Alzheimer's disease from a normal c omparison group. Second, to access this diagnostic information, it app ears that one must take advantage of the ability of PDP networks to de tect higher-order nonlinear relationships among the predictor variable s.