The vegetative state is a devastating condition where patients awaken from their coma (i.e., open their eyes) but fail to show any behavioural sign of conscious awareness. Locked-in syndrome patients also awaken from their coma and are unable to show any motor response to command (except for small eye movements or blinks) but recover full conscious awareness of self and environment. Bedside evaluation of residual cognitive function in coma survivors often is difficult because motor responses may be very limited or inconsistent. We here aimed to disentangle vegetative from "locked-in" patients by an automatic procedure based on machine learning using fluorodeoxyglucose PET data obtained in 37 healthy controls and in 13 patients in a vegetative state. Next, the trained machine was tested on brain scans obtained in 8 patients with locked-in syndrome. We used a sparse probabilistic Bayesian learning framework called "relevance vector machine" (RVM) to classify the scans. The trained RVM classifier, applied on an input scan, returns a probability value (p-value) of being in one class or the other, here being "conscious" or not. Training on the control and vegetative state groups was assessed with a leave-one-out cross-validation procedure, leading to 100% classification accuracy. When applied on the locked-in patients, all scans were classified as "conscious" with a mean p-value of .95 (min .85). In conclusion, even with this relatively limited data set, we could train a classifier distinguishing between normal consciousness (i.e., wakeful conscious awareness) and the vegetative state (i.e., wakeful unawareness). Cross-validation also indicated that the clinical classification and the one predicted by the automatic RVM classifier were in accordance. Moreover, when applied on a third group of "locked-in" consciously aware patients, they all had a strong probability of being similar to the normal controls, as expected. Therefore, RVM classification of cerebral metabolic images obtained in coma survivors could become a useful tool for the automated PET-based diagnosis of altered states of consciousness.
Copyright © 2010 Elsevier Inc. All rights reserved.