A Stacking Framework for Multi-Classification of Alzheimer's Disease Using Neuroimaging and Clinical Features.

Abstract

BACKGROUND Alzheimer's disease (AD) is a severe health problem. Challenges still remain in early diagnosis. OBJECTIVE The objective of this study was to build a Stacking framework for multi-classification of AD by a combination of neuroimaging and clinical features to improve the performance. METHODS The data we used were from the Alzheimer's Disease Neuroimaging Initiative database with a total of 493 subjects, including 125 normal control (NC), 121 early mild cognitive impairment, 109 late mild cognitive impairment (LMCI), and 138 AD. We selected structural magnetic resonance imaging (sMRI) features by voting strategy. The imaging features, demographic information, Mini-Mental State Examination, and Alzheimer's Disease Assessment Scale-Cognitive Subscale were combined together as classification features. We proposed a two-layer Stacking ensemble framework to classify four types of people. The first layer represented support vector machine, random forests, adaptive boosting, and gradient boosting decision tree; the second layer was a logistic regression classifier. Additionally, we analyzed performance of only sMRI feature and combined features and compared the proposed model with four base classifiers. RESULTS The Stacking model combined with sMRI and non-imaging features outshined four base classifiers with an average accuracy of 86.96% . Compared with using sMRI data alone, sMRI combined with non-imaging features significantly improved diagnostic accuracy, especially in NC versus LMCI and LMCI versus AD by 14.08% . CONCLUSION The Stacking framework we used can improve performance in diagnosis of AD using combined features.