Sleep in patients suffering from disorders of consciousness (DOC) provides clinically relevant information. A behavioural manifestation of sleep-wake rhythms allows to differentiate between coma and DOC. Furthermore, sleep patterns have been proposed to be indicative of the preserved residual brain functions in vegetative (VS) and minimally conscious patients (MCS). Apart from diagnostic utility, it is also postulated that patients' prognosis can be positively affected by sleep restoration and improvement of sleep quality. However, reliable and valid characterization of the unusual sleep patterns in this clinical group continues to be challenging for neuroscientists. The main difficulties are frequent artefacts in clinical recordings and absence of well-defined EEG changes. We address both of these issues employing Permutation Entropy (PE) together with Artificial Neural Networks (ANN). The first method measures the complexity of EEG, even if the signal is superposed with eye-related artefacts. The computation of PE for a recording with known sleep stages provides a quantified, nonlinear representation of sleep. The ANN-based classifier, in turn, allows to find the relation (mapping) between topographically distributed PE values and a corresponding sleep stage. The training of ANN aims to discover the relationships and fix them as the network parameters (weights). Such a trained classifier can be applied to an unknown set of data (e.g., independent healthy subject, DOC patient).