Sparse inverse covariance network-based modeling for mild cognitive impairment classification

Recent advances in neuroimaging techniques have provided great potential for studying mild cognitive impairments (MCI). Brain function connectivity networks constructed from resting-state functional magnetic resonance imaging (rs-fMRI), have been widely used for classifying MCI from normal controls (NC). In this paper, a sparse representation-based method called sparse inverse covariance estimation (SICE) constructs functional brain connectivity and distinguishes MCI patients from NC. One limitation of SICE methodology is its sensitivity to regularization parameters. To address this issue, a nested-cross validation framework was added to tune the regularization parameter through the machine learning process. After acquiring the optimal regularization parameter, a two-stage feature selection approach is performed on the brain network-based feature matrix to select the most discriminative features. Next, a linear support vector machine (SVM) classifier is trained for classification using these selected optimal features. The experiment results indicated a cross-validation accuracy of 91.89% with a sensitivity of 83.3%, and a specificity of 96%. The positive results illustrate the excellent diagnostic power of the SICE method. The proposed method found comparative differences between brain regions in MCI patients versus NC patients, which is consistent with findings in reported literatures.