Learning Brain Connectivity Sub-networks by Group- Constrained Sparse Inverse Covariance Estimation for Alzheimer's Disease Classification

Abstract

Background/Aims: Brain functional connectivity networks constructed from resting-state functional magnetic resonance imaging (rs-fMRI) have been widely used for classifying Alzheimer's disease (AD) from normal controls (NC). However, conventional correlation analysis methods only capture the pairwise information, which may not be capable of revealing an adequate and accurate functional connectivity relationship among brain regions in the whole brain. Additionally, the non-sparse connectivity networks commonly contain a large number of spurious or insignificant connections, which are inconsistent with the sparse connectivity of actual brain networks in nature and may deteriorate the classification performance of Alzheimer's disease.Methods: To address these problems, in this paper, a new classification framework is proposed by combining the Group-constrained topology structure detection with sparse inverse covariance estimation (SICE) method to build the functional brain sub-network for each brain region. Particularly, to tune the sensitive analysis of the regularized parameters in the SICE method, a nested leave-one-out cross-validation (LOOCV) method is adopted. Sparse functional connectivity networks are thus effectively constructed by using the optimal regularized parameters. Finally, a decision classification tree (DCT) classifier is trained for classifying AD from NC based on these optimal functional brain sub-networks. The convergence performance of our proposed method is furthermore evaluated by the trend of coefficient variation.Results: Experiment results indicate that a LOOCV classification accuracy of 81.82% with a sensitivity of 80.00%, and a specificity of 83.33% can be obtained by using the proposed method for the classification AD from NC, and outperforms the most state-of-the-art methods in terms of the classification accuracy. Additionally, the experiment results of the convergence performance further suggest that our proposed scheme has a high rate of convergence. Particularly, the abnormal brain regions and functional connections identified by our proposed framework are highly associated with the underpinning pathological mechanism of the AD, which are consistent with previous studies.Conclusion: These results have demonstrated the effectiveness of the proposed Group- constrained SICE method, and are capable of clinical value to the diagnosis of Alzheimer's disease.

Effective Diagnosis of Alzheimer’s Disease via Multimodal Fusion Analysis Framework

“…On the other hand, the multimodal random forest, based on fusion features, has more obvious advantages than the traditional single-modal method (Chanel et al, 2016; Guo et al, 2017). For instance, Rosa et al (2015) used sparse network-based models to identify brain disease patients with an accuracy rate of 79%, and Li et al (2018) employed group-constrained sparse inverse covariance which achieved about 80% accuracy in AD recognition. The average classification accuracy of the multimodal random forest is 83.33%.…”

Sparse inverse covariance matrix estimation via the $ \newcommand{\e}{{\rm e}} \ell_{0}$ -norm with Tikhonov regularization

“…Thus, Σ −1 defines a minimal I-map Markov network [24], whereby nonzero entries correspond to edges in the network. Following the parsimony principle, Σ −1 is always assumed to be sparse in many applications, such as biological inference in gene networks [1,41] and analysis of fMRI brain connectivity data [16,27]. Hence, sparsity regularized negative log-likelihood minimization becomes a popular approach for estimating the sparse inverse covariance matrix.…”

“…Sparse inverse covariance matrix estimation is a fundamental problem in constructing a Gaussian network model, which uses a graph-based representation as the basis for compactly encoding a multivariate Gaussian distribution over a high-dimensional space [24]. Its applications range from causal inference [2,19,29], biological inference in gene networks [1,41], and detection of brain connectivity [16,27]. The inverse covariance matrix Σ −1 of a Gaussian distribution, called information matrix (or precision matrix), captures independencies between pairs of variables, conditioned on all of the remaining variables in the model.…”