Abstract :
[en] Introduction: Functional connectivity has been successfully used to discriminate non-sedated patients with disorders of consciousness (Demertzi et al., 2015). However, on clinical demand, patients are evaluated under sedation to restrict motion, which considerably limits the classification of patients based on functional connectivity. It has been previously shown that changes of the frequency properties of spontaneous BOLD signal are of cognitive relevance even in sleeping neonates (Alcauter et al., 2015). We therefore aimed at exploring the automatic discrimination of sedated patients in the clinical entities of minimally consciousness state (MCS) and unresponsive wakefulness syndrome (UWS), based on the frequency profile of the BOLD signal.
Methods: Forty-four patients with MCS (n=26) or VS/UWS (n=18), based on the Coma Recovery Scale-Revised (CRS-R), were scanned on a 3T MRI scanner. Images of the whole brain were acquired with BOLD-sensitive sequences (300 volumes, TR=2s, TE=30ms, voxel size=3x3x3 mm3) and a T1 (TR=2.3s, TE=2.47ms, voxel size = 1x1x1.2 mm3). Sedative agents (propofol, sevoflurane, or a combination of both) were administered using the minimum necessary dose. 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, and regression of motion parameters, no temrporal filtering was applied. The average power density between 0.01 and 0.1 Hz (classic frequency band for resting state analyses) was estimated and divided by the total power density, for each voxel. Supervised classification of patients in MCS or UWS was explored with Support Vector Machine classifiier using stratified 5-fold cross-validation. The clusters with significant differences between groups (p<0.005, uncorrected; cluster size > 10 voxels) in the training sets were selected as features. The 5-fold validation was repeated 20 times to estimate the
variability of the classification accuracies and the frequency of each voxel being selected as a relevant feature.
Results:The average classification accuracy was 79%±5 (SD), with average sensitivity 76%±10, and specificity 81%±9. The most frequently selected regions as features included the superior parietal lobule (Frequency: 100%; MNI x, y, z (mm): -26, -50, 64), putamen (97%; -30, -6, -8), occipital fusiform gyrus (92%; -34, -70, -20), occipital pole (65%; 22, -98, 16), angular gyrus (54%; -60, -58, 32).
Conclusions: The power spectral density of the spontaneous BOLD signal under anesthesia allowed to classify individual patients with MCS and UWS with 79% accuracy. The most frequent selected features included association areas in the parietal and occipital lobes and the putamen. Further validation with independent cohorts is needed to generalize the current findings. Taken together, the use of power spectral density may represent an alternative to functional connectivity to classify patients with consciousness disorders under anesthesia, therefore capturing properties of conscious function beyond reportability.
