Alzheimer’s disease is a progressive neurodegenerative disorder which is characterized by the impairment in memory and disturbance in at least one other thinking function such as language or perception of reality. There is no a single test for the diagnosis of Alzheimer’s disease. A properly trained physician can diagnose the disease based on laboratory tests, genotyping, patient’s history, clinical observations, cognitive testing and brain imaging such as PET, MRI and fMRI. Functional Magnetic Resonance Images (fMRI) pinpoint dysfunctional areas of the brain and reveal a variety of differences between demented and healthy subjects even in the early stages of the disease.

The aim of the research work of the Unit of Medical Technology and Intelligent Information Systems is to exploit all possible features, that can be extracted from fMRI experiment and express AD related changes, in order to classify a person as healthy or AD and to classify the stages of the disease. An automated supervised method is proposed. The method consists of five stages (Fig. 1). First the fMRI data are preprocessed in order to remove confounding effects and make the data better meet the assumptions of the GLM which is used in the modeling of fMRI data (second stage of the method). In the third stage, features, which characterize AD patients, are extracted. A feature selection algorithm is applied to the set of the extracted features in order to identify and remove as much irrelevant and redundant information as possible (fourth stage). Finally, the Random Forests (RF) classification algorithm and some modifications of it are applied to distinguish AD and healthy subjects (Fig. 2). Classification has been made using majority and weighted voting schemes (Fig. 3).

The proposed method was evaluated using three different datasets. The first consists of two groups: healthy old and demented old subjects. The second dataset is the same as the first one but demented old subjects are divided in subjects with very mild and mild AD. The third dataset is the same as the second one but also includes healthy young subjects. In Table I the sensitivity and specificity that were produced in each case are reported. The 10-fold stratified cross validation procedure was used for the evaluation of the proposed method. The four versions of RF algorithm were also applied in order to classify AD progress. For this purpose the demented group was divided in those with very mild and mild dementia (three class problem) and the healthy young adults participated in the dataset (four class problem). The results of the classification of AD progress are reported in Table I.

TABLE I: CLASSIFICATION OF AD AND IT’S PROGRESS USINF ONLY fMRI FEATURES

The results reported in TABLE I was achieved using only fMRI features. The first three versions of Random Forests classification algorithm were tested using all extracted features. The achieved accuracy, sensitivity and specificity are reported in Table II.

TABLE II: CLASSIFICATION OF AD AND IT’S PROGRESS USINF ONLY ALL EXTRACTED FEATURES

TABLE III: COMPARISON WITH OTHER METHODS REPORTED IN THE LITERATURE