Diagnosis of Alzheimer’s Disease with Extreme Learning Machine on Whole-Brain Functional Connectivity

Abstract

The analysis of human brain fMRI subjects can research neuro-related diseases and explore the related rules of human brain activity. In this paper, we proposed an algorithm framework to analyze the functional connectivity network of the whole brain and to distinguish Alzheimer’s disease (AD), mild cognitive impairment (MCI), and cognitively normal (CN). In other studies, they use algorithms to select features or extract abstract features, or even manually select features based on prior information. Then, a classifier is constructed to classify the selected features. We designed a concise algorithm framework that uses whole-brain functional connectivity for classification without feature selection. The algorithm framework is a two-hidden-layer neural network based on extreme learning machine (ELM), which overcomes the instability of classical ELM in high-dimensional data scenarios. We use this method to conduct experiments for AD, MCI, and CN data and perform 10-fold cross-validation. We found that it has several advantages: (1) the proposed method has excellent classification accuracy with high speed. The classification accuracy of AD vs. CN is 96.85%, and the accuracy of MCI vs. CN is 95.05%. Their AUC (area under curve) of ROC (receiver operating characteristic curve) reached 0.9891 and 0.9888, respectively. Their sensitivities are 97.1% and 94.7%, and specificities are 96.3% and 95.3%, respectively. (2) Compared with other studies, the proposed method is concise. Construction of a two-hidden-layer neural network is to learn features of the whole brain for the diagnosis of AD and MCI, without the feature screening. It avoids the negative effects of feature screening by algorithm or prior information. (3) The proposed method is suitable for small sample and high-dimensional data. It meets the requirements of medical image analysis. In other studies, its classifiers usually deal with several to dozens of feature dimensions. The proposed method deals with 4005 feature dimensions.