Meta-ensembles of Classifiers for Alzheimer’s Disease Detection Using Independent ROI Features

Abstract

Due to its growing social impact, prodromal detection of Alzheimer’s disease is of paramount importance. Biomarkers based on Magnetic Resonance Imaging (MRI) are one of the most sought results in the neuroscience community. In this paper we evaluate several ensembles of classifiers trained and tested in a two level ensemble scheme as follows: the 116 regions of interest (ROI) of the Anatomical Automatic Labeling (AAL) brain atlas are used to compute disjoint feature sets from the Grey-matter probability maps from the segmentation of the T1 weighted MRI of each subject; ROI features are the summary statistics inside this ROI; one ensemble of classifiers is trained on each independent ROI feature data set; the final classification of each subject is given by the combination of the classifications of each ROI, as meta-ensemble classifier. Experiments are performed on the 416 subjects (316 controls and 100 patients) of the OASIS database. We perform a hold-out of the 20% of the data for model selection, computing a leave-one-out validation on the 80% remaining data. Results are computed without circularity. Tested classifiers are the Extreme Learning Machines (ELM), Bootstrapped Dendritic Computing (BDC), Hybrid Extreme Random Forest (HERF) and Random Forest (RF). The best performance achieved is 80.8% accuracy, 77.1% specificity and 92.5% specificity with BDC. We also report the most discriminant ROIs obtained in the model selection phase.