Sparse Discriminative Feature Selection for Multi-class Alzheimer’s Disease Classification

In neuroimaging studies, high dimensionality and small sample size have been always an issue, and it is common to apply a dimension reduction method to avoid the over-fitting problem. Broadly, there are two different approaches in reducing the feature dimensionality: feature selection and subspace learning. When it comes to the feature interpretability, the feature selection approach such as the sparse regularized linear regression method is preferable to the subspace learning methods, especially in Alzheimer’s Disease (AD) diagnosis. However, based on recent machine learning researches, the subspace learning methods presented promising results in various applications. To this end, in this work, we propose a novel method for discriminative feature selection by combining two conceptually different methodologies of feature selection and subspace learning in a unified framework. Specifically, we integrate the ideas of Fisher’s linear discriminant analysis and locality preserving projection, which consider, respectively, the global and local information inherent in observations, in a regularized least square regression model. With the help of global and local information in data, we select class-discriminative and noise-resistant features that thus help enhance classification performance. Furthermore, unlike the previous methods that mostly considered only a binary classification, in this paper, we consider a multi-class classification problem in AD diagnosis. Our experiments on the Alzheimer’s Disease Neuroimaging Initiative dataset showed the efficacy of the proposed method by enhancing the performances in multi-class AD classification.