Predictive models of minimal hepatic encephalopathy for cirrhotic patients based on large-scale brain intrinsic connectivity networks.

We aimed to find the most representative connectivity patterns for minimal hepatic encephalopathy (MHE) using large-scale intrinsic connectivity networks (ICNs) and machine learning methods. Resting-state fMRI was administered to 33 cirrhotic patients with MHE and 43 cirrhotic patients without MHE (NMHE). The connectivity maps of 20 ICNs for each participant were obtained by dual regression. A Bayesian machine learning technique, called Graphical Model-based Multivariate Analysis, was applied to determine ICN regions that characterized group differences. The most representative ICNs were evaluated by the performance of three machine learning methods (support vector machines (SVMs), multilayer perceptrons (MLP), and C4.5). The clinical significance of these potential biomarkers was further tested. The temporal lobe network (TLN), and subcortical network (SCN), and sensorimotor network (SMN) were selected as representative ICNs. The distinct functional integration patterns of the representative ICNs were significantly correlated with behavior criteria and Child-Pugh scores. Our findings suggest the representative ICNs based on GAMMA can distinguish MHE from NMHE and provide supplementary information to current MHE diagnostic criteria.