Abstract :
[en] Introduction: We have previously used machine learning to separate minimally conscious from unconscious patients based on network-level resting state fMRI connectivity (Demertzi et al., 2015). By means of a Support Vector Machine (SVM) classifier and using connectivity values extracted from the "auditory network" as features, we identified 20/22 patients leading to an accuracy of 91% in differentiating the two clinical populations when validated on an independent dataset. The features of the "auditory network" encompassed temporal and occipital regions, raising the question whether the classifier captures merely sensory information (i.e. auditory, visual) or higher-order cortical organization pertaining to conscious conditions. To that aim, we performed generalization trials on data collected from congenitally deaf and congenitally blind patients with preserved global states of consciousness. We hypothesized that if the classifier captures global states of consciousness, the new tested subjects would belong to the class of the minimally conscious patients.
Methods: The training dataset included 45 patients with disorders of consciousness (26 MCS, 19 UWS). The generalisation dataset included nine congenitally blind and eight congenitally deaf subjects. Functional images (300 for the training and 280 for the generalisation dataset) and a structural image were acquired on 3T MRI systems. Preprocessing of functional images included slice-time correction, realignment, segmentation, normalisation, and smoothing (6mm FWHM). Noise reduction included detection and regression of motion outliers (ART toolbox), anatomical component-based correction, regression of motion parameters, and band-pass filtering (0.008-0.09Hz). Structural images were reoriented, spatially normalized to MNI space, segmented and smoothed (6mm FWHM). For all preprocessing steps we used SPM8 (Ashburner et al, 2005). Seven spherical regions were used as seeds (Maudoux et al., 2012) from which averaged time-series were calculated to estimate whole brain correlation r-maps. R-maps were then converted to normally distributed Fisher's z transformed correlation maps. One-sample t-test was ordered to estimate seed-tovoxel whole brain connectivity. Three connectivity values per subject were extracted using binary masks, which were calculated on patients with disorders of consciousness and which encompassed bilateral temporal and occipital areas. A linear SVM with default regularization parameter C=1 was trained on the data (45X3 features) from patients with disorders of consciousness.
Results: Out of the nine congenitally blind subjects, one was placed in the class of unconscious patients. Out of the 8 congenitally deaf subjects, one was placed in the class of unconscious patients.
Conclusions: Our results indicate that joint temporo-occipital functional connectivity exceeds mere sensory perception and it reflects higher-order functional organisation indicative of preserved conscious states. These findings assist to better comprehend aspects of conscious processes, which are further informative to evaluate covert cognition in noncommunicating conditions.
