Machine Learning in the Diagnosis of Disorders of Consciousness: Opportunities and Challenges

The detection of conscious awareness in patients with disorders of consciousness using behavioral signs is challenging in the presence of sensory, motor, or executive function deficits. Brain signal measurements provide additional information for patient stratification, and the large amounts and multidimentional nature of the obtained data motivate the application of machine learning (ML) methods. We review some of the recent studies applying this approach to data from electroencephalography, diffusion magnetic resonance imaging (MRI), and resting-state functional MRI. Each of these modalities provided features correlating with the behaviorally determined levels of consciousness. At the same time, considering the accuracy of the obtained classification algorithms, we highlight the problem of simultaneous model selection and accuracy estimation using the same sample, which requires special techniques of statistical analysis, otherwise the obtained accuracy estimates can be biased and prone to misinterpretation. Additionally, there is a need for greater generalization ability of the classifiers over the whole spectrum from the UWS to the conscious state. The use of ML within brain-computer interfaces for the detection of command following is discussed as a promising complementary method able to detect patients with cognitive motor dissociation, and its results may prove useful as labels in the training of algorithms for the direct stratification of DOC patients.