Abstract: Previously proposed sparse logistic regression algorithms cannot make the most of the spatial structure information of brain, so it is not appropriate to be applied directly in whole-brain fMRI data classification. To address the problem, two penalty terms of characterizing sparsity and spatial structure are simultaneously introduced into traditional logistic regression algorithm, which produces a generalized sparse logistic regression algorithm. Specifically, a flexible penalty term that characterizes the spatial structure is designed. By tuning parameters of this penalty term, the spatial structure information of brain can be fully utilized. Under the framework of minorization-maximization, an iterative process is designed to solve the corresponding optimization problem, which guarantees that an explicit solution can be obtained in each iteration and a locally optimal solution for the problem can be found. Experimental results show that the proposed algorithm outperforms existing sparse logistic regression algorithms in classification accuracy.