Predicting Alzheimer's Disease Based on Network Topological Latent Representations

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative disease. Right now there is no cure for AD. As we know, the brain structure is constantly changing as it progresses to AD. If we can detect these changes, it may be helpful to discover new biomarkers and early diagnosis of AD. We use the deep learning method DeepWalk to extract the structural features of each subject based on brain function connection. The method takes a brain topology network based on functional connection transformations as input and outputs a potential feature representation for each brain region in the brain network structure. In the current research, a new potential biomarker is found, which is based on a structural feature that can represent changes in brain connectivity. In our experiment, 79 Resting-state fMRI were used from ADNI dataset, including 49 normal controls (NCs) and 30 AD patients. By analyzing the structural features between AD patients and NCs, obvious differences are detected in the temporal lobe, parietal lobe, and frontal lobe. We use a support vector machine (SVM) model to evaluate the performance of these structural features, and its classification performance achieves 80.36% accuracy (specificity = 73.67%, sensitivity = 87.17%). In general, these findings indicate that structural features may be a new potential biomarker between NCs and AD patients.