The question of whether someone is conscious of themselves and their environment may initially appear to be a philosophical thought experiment. However, for many families and clinicians throughout the world, it is a challenge with profound implications for decision-making after severe brain injury. The current standard approach to answering the question of consciousness in the clinic is to observe the patient’s behaviour for evidence of purposeful action, such as tracking a moving object. However, because of inequalities in healthcare provision both between and within countries, many patients never receive standardized behavioural assessment. Consequently, misdiagnoses of so-called prolonged disorders of consciousness (PDOC) occur at alarmingly high rates: ~40% when standardized assessment is not available and ~30% when a standardized assessment is not conducted longitudinally (Wannez et al., 2017). With the recent confirmation from the UK’s Supreme Court that withdrawal of clinically assisted nutrition and hydration from individuals with PDOC may occur without the involvement of the courts (UK Supreme Court, 2018), access to accurate diagnostic tools is all the more imperative. In this issue of Brain, Engemann and co-workers present results from a trans-European collaboration that offer hope for accurate, accessible, and objective diagnoses of consciousness in this challenging patient group (Engemann et al., 2018).
Over the past decade, researchers have identified an array of markers within the EEG of patients with PDOC that differ between those who are conscious (i.e. the minimally conscious state) and those who are entirely unconscious [i.e. the vegetative state, also known as unresponsive wakefulness syndrome (UWS)]. These markers may be task-based changes in the EEG that occur in response to stimuli, such as sounds or verbal instructions, or task-free features of the patient’s EEG at rest. Engemann et al. (2018) apply advances in machine learning to the largest ever dataset of PDOC brain data and demonstrate that a combination of theoretically- and empirically-motivated EEG markers can accurately diagnose patients’ levels of consciousness.
To test the efficacy of a machine learning method, researchers first require a set of training data for which they and their algorithm know the ‘truth’—i.e. which data correspond to patients who are conscious by standardized assessment, and which belong to those who are not conscious. The researchers then apply the trained model to previously unseen test data from one patient in order to estimate their ‘true’ diagnosis. Crucially, the model described by Engemann et al. (2018) accurately diagnoses patients even when the training and test data were recorded in different countries, with different EEG equipment, and different EEG protocols. The success of this model, therefore, creates the real possibility for an objective online tool that families and clinicians across the world could use with locally recorded EEG data, and that could consequently improve diagnostic accuracy and subsequent decision-making. Indeed, through a range of stress tests, the authors demonstrate that their model performs well with only a few minutes of data recorded from 16 electrodes; fewer than the 19-channel EEG montage typically available in clinical EEG protocols.
While the diagnostic accuracy of the model is impressive, there is nevertheless potential for the model to misdiagnose patients due to fluctuations in levels of consciousness over time. Indeed, clinicians must conduct multiple behavioural assessments of consciousness before they can achieve a stable and accurate diagnosis (Wannez et al., 2017). Equally, the level of consciousness that is evident in a patient’s EEG is likely to fluctuate both within and across days, and may therefore require multiple EEG sessions before a stable diagnosis can be reached. Importantly, as the diagnostic model described by Engemann et al., 2018 is robust with brief EEG recordings from a range of EEG systems and protocols, these necessarily longitudinal assessments of EEG may be more tractable now than ever before.
The authors report that power in the alpha band of the EEG (8–12 Hz) was consistently the most informative feature used by the models for accurate diagnosis. Alpha oscillations are the most prominent rhythm in the healthy brain and have long been associated with neuronal idling, such that high alpha power was thought to reflect a state of underlying cortical inactivity. While the alpha rhythm can be seen with the naked eye in the EEG of healthy individuals, it is considerably reduced in PDOC (Figure 1). Resting state alpha oscillations in the healthy brain are thought to be generated within the thalamus and its connections with cortex (Roux et al., 2013). Thalamic damage is perhaps the most consistent structural impairment observed across PDOC, and differentiates minimally conscious state from unresponsive wakefulness syndrome (Fernandez-Espejo et al., 2010). Therefore, EEG alpha power may be the most informative marker in this diagnostic model precisely because it reflects the relative level of damage to brain structures that are known to underlie a patient’s level of consciousness.
The role of alpha oscillations in perception. The thalamus (highlighted in pink) is highly connected to the cortex and is implicated in generating alpha oscillations through thalamo-cortical loops. A schematic alpha oscillation (top right) demonstrates the proposed active inhibition mechanism of boosting (green rectangle) or inhibiting (red rectangle) sensory processing by means of power changes, and the perceptual boost observed at specific phases of the alpha oscillation. The power spectrum (bottom) of a healthy individual (black line) shows a clear peak in the alpha band (at 10 Hz), which is entirely absent from a typical patient with a diagnosis of unresponsive wakefulness syndrome [grey line; data taken from Beukema et al. (2016); doi:10.1016/j.nicl.2016.08.003].
A complementary interpretation is that EEG alpha power also reflects an important aspect of the capacity for contents of consciousness in PDOC. Indeed, recent evidence suggests that alpha oscillations play a functional role in both cognition and perception. For example, visual perception relies on alpha oscillations to segment incoming visual information into bite-sized snapshots of ~100 ms for visual cortex to process (VanRullen, 2016). When we pay attention to one side of our visual field, alpha power decreases over contralateral visual cortex and increases over ipsilateral visual cortex. The active inhibition hypothesis of alpha oscillations (Jensen and Mazaheri, 2010) proposes that increases in alpha power reflect inhibition of processing in task-irrelevant brain regions, while decreases in alpha power provide a boost to neural processing in task-relevant regions. Furthermore, in support of a functional role of alpha oscillations, the extent to which alpha becomes lateralized during spatial attention correlates with detection accuracy of visual targets (see Jensen and Mazaheri, 2010 for a review). Therefore, without a functional alpha rhythm to orchestrate the activity of relevant cortical regions, PDOC cortex becomes incapable of functional perception and cognition. Patients would not perceive the world in discrete perceptual cycles and would be unable to selectively attend to sensory stimuli, such as a relative’s voice or the instructions of a rehabilitation specialist. While the alpha power markers used in the diagnostic model described by Engemann et al. (2018) were summarized across the whole head, recent evidence of multiple, functionally dissociable neural sources of alpha oscillations (Sokoliuk et al., 2018) suggests that a more fine-grained picture of the relative preservation of both levels and contents of consciousness in PDOC may be possible.
The importance of resting alpha power in this diagnostic model also raises the question of whether it is futile to pursue task-based, or ‘active’, methods for detecting consciousness. Perhaps the most well-known task-based paradigm involves asking patients to imagine that they are playing tennis and looking for activity in task-appropriate regions of their brain in order to infer that they obeyed the instruction, and are therefore conscious (Owen et al., 2006). Evidence of consciousness from such ‘active’ methods is certainly compelling and appeals to our willingness to consider that someone is conscious only if they can show us with their (physical or mental) behaviour. However, a recent large-scale study of PET and functional MRI data showed that active paradigms were less useful for diagnosis than simple task-free measures (Stender et al., 2014). Furthermore, in the study described in this issue of Brain, task-based markers of consciousness contributed little to the accuracy of diagnoses compared with task-free, passive, markers. Nevertheless, in our opinion, both task-free and task-based data are vital for solving the clinical challenges posed by PDOC. While task-free data may provide evidence for the preservation of a set of brain functions and structures that are minimally required for demonstrating consciousness through overt or covert behaviour, they cannot answer a host of important questions about the patient’s cognitive abilities. For example, can the patient understand what is being said to them by their families? Can the patient maintain attention for sufficient time for rehabilitation efforts to succeed? Can the patient form or retrieve memories of important events? The answers to these questions may arguably assist families more in their decision-making than abstract questions of whether someone is ‘conscious’. It may be that, in future, the method described in the current study will allow accurate diagnosis of the global state of consciousness of a given patient, while subsequent task-based examinations can outline the cognitive profile of the patient to help guide decision-making and rehabilitation efforts.
Finally, the study by Engemann et al. (2018) is a landmark in open-source big data in the field of PDOC. Indeed, the authors performed all analyses in the study with the open-source programming language of Python, and made the machine learning ‘recipe’ publicly available online alongside the scripts for EEG feature extraction. For non-behavioural measures of consciousness to achieve a true impact on clinical practice, the field must continue to generate and share larger and larger open access datasets, thus developing more accurate methods to assist families and clinicians with their most difficult decisions.