Correlation between Cognitive Function Scores and the Response of a Neural Network Classiier for Spect Data in Patients with Alzheimer's Disease

An artiicial neural network (ANN) was used to analyse and classify 99m Tc]-d,l-HMPAO SPECT data sets from 25 patients with a diagnosis of probable Alzheimer's disease and 25 healthy control subjects. Methods: Prior to the ANN analysis, the original set of 8 semiquantitative CBF values from cortical ROIs representing each subject was transformed into a new set of 8 values consisting of sums and numerical diierences between corresponding left and right regions. An ANN was then trained on the binary classiication problem based on these values. During training the ANN was pruned in order to discover the optimal ANN architecture. After training, a calculation of the saliencies of the ANN parameters ranked the lobar regions according to their importance in the classii-cation task. The continuous output of the ANN was compared to the corresponding Mini-Mental State Examination (MMSE) scores in order to see whether the ANN output could be interpreted as a measure of disease severity albeit the ANN was trained on the binary classiication task only. Results: The classiication problem was found to be linearly separable. The resulting ROC curve of the linear ANN classiier had a ROC area equal to 0.93. A study of the saliencies of the ANN parameters showed that the response of the ANN was largely based on 4 out of the 8 input variables. The correlation coeecient between the response 3 of the linear ANN classiier and the MMSE-scores was-0.7 (p < 0.001). Conclusion: By using the method of pruning the classiication task has been found to be linearly separable. The ROC area of the linear ANN classiier was comparable to that of more complex ANN classiiers found in similar analyses. The concept of parameter saliencies may provide important information about the involved regions. Although trained on the binary classiication problem only, the signiicant correlation between MMSE scores and the corresponding continuous output of the linear ANN indicates that the ANN output provides information about the disease severity. Thus, the MMSE scores reeect in a natural (linear) way the changes in rCBF values.