Meeting Banner
Abstract #1000

Alzheimer's Disease Prediction Based on Machine Learning Methods Applied to Multimodal MR Features

Giovanni Giulietti1, Michael Dayan1, Laura Serra1, Elisa Tuzzi1, Barbara Spano'1, Mara Cercignani2, Carlo Caltagirone3, 4, Marco Bozzali1

1Neuroimaging Laboratory, Santa Lucia Foundation, Rome, Italy; 2Clinical Imaging Sciences Centre, Brighton & Sussex Medical School, Brighton, United Kingdom; 3Clinical and Behavioural Neurology, IRCCS Santa Lucia Foundation, Rome, Italy; 4Departement of Neuroscience, "Tor Vergata" University, Rome, Italy

In the current study, we investigated the classification between healthy subjects and patients with Alzheimers disease, using structural (T1) and diffusion (DWI) MR data as input to Support Vector Machine (SVM) classifiers. SVM based on T1 features had higher discrimination capability relative to SVM based on DWI, but the best classification performance (92.6% of accuracy) was obtained combining them. We achieved satisfactory result despite the utilization of a small number of features, considering that it is not uncommon to use hundreds features to improve the classification performance. This evidence make our approach suitable to be adopted into clinical practice.

Abstract PDF Poster PDF

Keywords

according account accuracy acquisition additional allowed among anisotropy applied basis belong best brain builds built category characteristic classification classified classifier classifiers classify clinical collecting collinear computation compute computed constant contemporary correctly create cross custom dataset dementia demographic despite diagnosed diffusion diffusions diffusivity direct discriminate disease even every examination example examples expected exploiting expressed feature features field finally finding focuses formed former foundation fraction fractional furthermore grid growing healthy impaired improve included indicates intelligent inverse investigated isotropic kernel last learning leave library linear machine magnetization maps margin matched medical mental misclassified model multiplying native nearly neurology nominal observation observations optimized original patient patients performance planar population practice prediction proceedings processed proportion recruited remaining repeating resting sample school sciences script search segmentation sensitivity separately session seven smoothing space spatial specificity stages status step structural subject subjects suitable supervised support systems thick thousands threshold tissues together training transactions transformations underwent vector volume whole whose