Accuracy and generalization capability of an automatic method for the detection of typical brain hypometabolism in prodromal Alzheimer disease.

Abstract

The aim of this study was to verify the reliability and generalizability of an automatic tool for the detection of Alzheimer-related hypometabolic pattern based on a Support-Vector-Machine (SVM) model analyzing 18F-fluorodeoxyglucose (FDG) PET data. The SVM model processed metabolic data from anatomical volumes of interest also considering interhemispheric asymmetries. It was trained on a homogeneous dataset from a memory clinic center and tested on an independent multicentric dataset drawn from the Alzheimer's Disease Neuroimaging Initiative. Subjects were included in the study and classified based on a diagnosis confirmed after an adequate follow-up time. The accuracy of the discrimination between patients with Alzheimer Disease (AD), in either prodromal or dementia stage, and normal aging subjects was 95.8%, after cross-validation, in the training set. The accuracy of the same model in the testing set was 86.5%. The role of the two datasets was then reversed, and the accuracy was 89.8% in the multicentric training set and 88.0% in the monocentric testing set. The classification rate was also evaluated in different subgroups, including non-converter mild cognitive impairment (MCI) patients, subjects with MCI reverted to normal conditions and subjects with non-confirmed memory concern. The percent of pattern detections increased from 77% in early prodromal AD to 91% in AD dementia, while it was about 10% for healthy controls and non-AD patients. The present findings show a good level of reproducibility and generalizability of a model for detecting the hypometabolic pattern in AD and confirm the accuracy of FDG-PET in Alzheimer disease.