Gradient self-weighting linear collaborative discriminant regression classification for human cognitive states classification

Abstract

In recent decades, huge volumes of data are available to inspect human brain activities for disease detection. Specifically, the functional magnetic resonance imaging (fMRI) is a powerful tool to enquire the brain functions. In fMRI, identifying the active patterns of the specific cognitive state is one of the emerging concerns for neuroscientists. The high-dimensional features make fMRI data difficult for mining and classification, because if the volume of the data space increases, then the acquired data become sparse, which leads to the “curse of dimensionality” problem. To address this concern, a new feature selection and classification methodology was proposed for classifying the human cognitive states from fMRI data. Initially, the fMRI data were collected from the StarPlus and Haxby datasets. Then, k-nearest neighbors algorithm (k-NN)-based genetic algorithm was developed to choose the optimal voxels from the active region of interests. The proposed approach selects the data to feature subsets based on k-NN algorithm, so the data volume was effectively reduced and the voxel information was maintained significantly. The most informative voxels were given as the input for gradient self-weighting that produces an optimal weight value. Respective weight value was added to the projection matrix of linear collaborative discriminant regression classification for identifying the future projection matrix that reduces the error between two individual voxels in subspace. The experimental outcome shows that the proposed methodology improved the accuracy in fMRI data classification up to 0.7–23% compared to the existing methods.