Multimodal study of Alzheimer's Disease using machine learning methods

Abstract

It is estimated that there are around 47 million people affected by Alzheimer's disease (AD) in the world. For this reason, there are more and more studies and data collection about groups of patients, using techniques such as magnetic resonance imaging (MRI) that provide information about the evolution of AD. But not only the MRI are the significant features for the diagnosis of AD. Others biomarkers and data such as Positron-emission tomography (PET), APOE genotype, the age of the patient, or even the years that these patients have dedicated to education, can provide complementary information for diagnosis. In this study, we present a system combining these different types of data using machine learning techniques to predict not only if a patient has AD but also if a patient is in the prodromal stage, the mild cognitive impairment (MCI). To develop a predictive training classifier, we have used different clinical data and several types of biomarkers extracted from MRI, such as cortical thickness or volumetric measures from different different regions of brain from 813 patients. Then, combining these data we built differents classifiers with two methods of machine learning, support vector machine and decision tree. Using the first method we achieved an accuracy of 0.94 when classifying between subjects with AD and normal cognitive (CN), and an accuracy of 0.66 when distinguishing between AD, MCI and CN. Using decision tree we achieved an accuracy of 0.69 in binary classification (MCI-CN).