[P3–385]: VISUAL READING OF AMYLOID‐PET IN MCI CHALLENGED: SHOULD WE CONSIDER ALTERNATIVE METHODS?

healthy controls (HC) and patients with amnestic mild cognitive impairment (MCI) were dichotomized into groups of low amyloid (Aß-, HC: n 1⁄4 165, MCI: n 1⁄4188) and high amyloid status (Aß+, HC: n 1⁄4 73, MCI: n 1⁄4 231), with AD dementia group (n 1⁄4 98). Grey matter volume and FDG-PET were obtained within 90 ROIs of the Desikan-Killiany atlas. Feature selection and classification was based on a combination of inner and outer-loop 10-fold cross-validation scheme. Within each fold of a 10-fold cross-validation (outer-loop), feature selection based on the Information Gain criterion was determined within an inner-loop 10-fold cross-validation. The most discriminative features were then tested in the unseen test sample of the outer-loop fold. Overall classification accuracy via support vector machine was calculated as the average outerloop cross-validation accuracy. Results:Classification accuracy for the discrimination between AD and HC Aßwas > 90% for any single modality, and did not significantly benefit from a combination of modalities. For predicting conversion from MCI to AD (median follow-up time 1⁄4 36 months), the highest classification accuracy was 80.4%, when applying a combination of FDG-PET (for best features see Figure) and neuropsychological variables (world recall, Benton naming, tests of executive function). For discriminating Aß status in HC or MCI as well as conversion from HC to MCI, we observed classification accuracies between 60 – 70%. Conclusions:The combination of FDG-PET and neuropsychological variables produced clinically useful classification accuracy to predict conversion of AD dementia at MCI stage, but remained still below clinical utility for the detection of preclinical AD.