Cogito ergo sum by Imaging: The utility of neuroimaging in assessing state of consciousness

Cogito ergo sum by Imaging

The  utility  of  neuroimaging  in  assessing  state  of  

consciousness  

MSc  in  Brain  and  Cognitive  Sciences   (Cognitive  Neuroscience)   University  of  Amsterdam  

Table of Contents

Abstract

p. 2

Introduction

p. 3-8

Passive Paradigms

p. 9-14

Active Paradigms

p. 14-21

Discussion

p. 21-25

Reference Section

p. 26-31

Abstract

Unresponsive  wakefulness  syndrome  (UWS)  classifies  patients  that  are  completely  unaware   of   themselves   and   the   environment   yet   behaviorally   display   sleep-­‐wake   cycles.   With   the   advancement  of  modern  medicine,  the  prevalence  of  this  condition  has  increased,  giving  rise   for  the  need  of  an  accurate  diagnostic  tool  to  assess  awareness.  Assessing  awareness  within   this  group  of  patients  is  a  particularly  difficult  due  to  the  heterogeneity  of  injury  profiles,   which   in   turn   gives   rise   to   a   patient   population   with   a   wide   variety   of   different   cognitive   deficits.  Behavioral  assessment  is  the  most  common  method  to  determine  the  presence  of   awareness,  yet  it  relies  on  the  ability  of  the  subject  to  behaviorally  respond  in  a  consistent   manner.  Recently  neuroimaging  methods  have  been  developed  specifically  to  address  this   issue.  Covert  awareness  can  be  assessed  through  passive  (paradigms  simply  measuring  the   subject’s  neural  response  to  a  stimuli)  or  active  (paradigms  instructing  patients  to  willfully   modulate  their  neural  activity  upon  command)  neuroimaging  methods.  The  author  of  this   review   compares   and   contrasts   the   different   neuroimaging   approaches   to   assess   covert   awareness,  while  focusing  on  the  most  promising  methods.  Finally  concluding  by  outlining  a   multidimensional  awareness  evaluation  approach  that  is  viable  for  clinical  implementation.                                                

Introduction

Disorders of consciousness (DoC) is a diagnostic umbrella term for patients in differing conscious states. This range of disorders include brain death, coma, vegetative state and minimally conscious state (Bernat, 2006). DoC are a relatively new phenomenon, resulting from unintended consequences brought upon by advances in internal medicine (Monti, 2012). One advance of particular interest is the artificial respirator, since it allows patients to continue living without sustained breathing. Occasionally the injury sustained is too severe for a full recovery, thus resulting in a coma. The three main outcomes of a coma are either recovery, death or vegetation. A vegetative diagnosis is impingent upon the reemergence of spontaneous eye-opening (Giacino, Fins, Laureys, & Schiff, 2014; Gosseries, Zasler, & Laureys, 2014). Vegetative state (previously called apallic syndrome) is defined as a “condition of complete unawareness of the self and the environment, accompanied by sleep-wake cycles, with either complete or partial preservation of hypothalamic and brain-stem autonomic functions.” (Ashwal et al., 1994, p. 1500). This diagnosis has a wide variety of etiologies, yet the two main causes are head trauma (TBI) or hypoxic-ischemic encephalopathy (Ashwal et al., 1994). These two main causes can lead to substantially different neuronal injury profiles. Although diffuse neuronal damage is prevalent in both, TBI also causes severe focal damage. This difference in injury profiles will be of particular interest later in the manuscript when we discuss the utility of neuroimaging to assess the presence or absence of consciousness, since etiology will effect which cognitive functions remain. In 2010, the European task force on disorders of consciousness decided to change the name of this syndrome to unresponsive wakefulness syndrome (UWS) (Laureys et al., 2010). It is the opinion of the author that UWS is a more suitable term since it does not presume a lack of awareness. This review will proceed to substitute vegetative state with unresponsive wakefulness syndrome (UWS), yet the term vegetative state (i.e. VS) will be present in direct quotes and figures from articles. Any reference of vegetative state should be taken as synonymous to UWS.

Recovery from UWS commonly leads to a minimally conscious state (MCS) diagnosis, which labels patients with fluctuating, yet reproducible behavioral responses. Similar to UWS these patients are wakeful, reflected by spontaneous eye-opening (Monti, Laureys, & Owen, 2010). Differential diagnosis, between UWS and MCS, has substantial implications since the main diagnostic difference is awareness (i.e. consciousness). Figure 1 illustrates the relation between UWS and MCS on a variety of dimensions: motor behavior production (y-axis), wakefulness (x-axis) and conscious awareness (z-axis). The shaded blue population of patients are behaviorally indistinguishable, yet can be either UW or MC. Since there is no direct way of assessing consciousness, diagnosis must rely on indirect measures, such as behavioral reports. This leads to this shaded blue population receiving an UWS diagnosis, regardless of their actual levels of awareness. Therein lies the problem with diagnosing UWS, leading to an unacceptably high incorrect classification of around 40% (Andrews, Murphy, Munday, & Littlewood, 1996; Bodien & Giacino, 2016; Gill-Thwaites & Munday, 2004; Schnakers et al., 2009). This incorrect classification rate reflects patients which hold a UWS diagnoses yet their levels of awareness are more indicative of a MCS diagnosis.

The goal of this review is to assess the diagnostic efficacy of neuroimaging tools in DoC. Specifically focusing on the ability of neuroimaging to classify the behaviorally unresponsive MCS patients (represented in the figure as shaded purple MCS). This distinction of awareness is crucial, as it has ethical and care-based implications, such as the withdrawal of life support (Syd & Johnson, 2016).

Figure 1 Reproduced from Monti (2012) under the creative commons license. The blue highlight

reflects the focus of this review: Behaviorally indistinguishable patients with varying levels of awareness.

The lack of a direct method to assess the absence of consciousness, necessary for an UWS diagnosis, leads to a logical fallacy. Formulated by Monti and colleagues (2010) as the “absence of evidence (of awareness) is taken as evidence of absence (of awareness)” (p. 295). This approach relies on a null result and in turn will always incorrectly classify a MC behaviorally unresponsive patient as UWS. Two types of variance can incorrectly confirm this logical fallacy; 1) prognostic error and 2) patient-specific error. Prognostic error can result from subjective bias by the clinician, which is impingent upon the range of behaviors sampled, the frequency of assessments and the behavioral diagnostic measure (Giacino et al., 2014). The most widely used behavioral measure for diagnosing DoC is the Glasgow coma scale (GCS) (Majerus, Gill-Thwaites, Andrews, & Laureys, 2005). The scale measures a subject’s behavioral response to external stimuli in three components: eye, verbal and motor. These components reflect the different modalities in which a subject could respond to a stimulus. A score of one on the GCS would reflect no behavioral response in any of the categories, while the upper bound (15) reflects a fully aware subject. The GCS has limitations which contribute to the high rate of misdiagnosis in DoC. One of the most prominent criticism is its failure to

the JFK Coma Recovery Scale-Revised (CRS-R), which is of particular relevance since it allows for an ‘uncertain diagnosis’ between UWS and MCS.

The second type of variance is patient specific error, a substantial amount results from underlying sensory defects and fluctuating levels of vigilance (Majerus, Bruno, Schnakers, Giacino, & Laureys, 2009; Owen, Schiff, & Laureys, 2009). The ability to respond is even further compounded by environmental factors such as sedative and analgesic mediation (Majerus et al., 2005). Neuroimaging tools can slightly mediate sensory defects as a source of variance since they circumvent the need for overt behavior. Some studies have been able to correctly diagnose behaviorally unresponsive MCS patients. These advances have led to the proposal to divide MCS into MCS+ and MCS-, reflecting differing levels of behavioral response (Bruno, Vanhaudenhuyse, Thibaut, Moonen, & Laureys, 2011). Yet this approach is futile, since there is a growing consensus in the field that awareness is part of a continuum, as represented in figure 1 (Monti, 2012; Monti, Laureys, et al., 2010; Sinai, Owen, & Naci, 2017). For convenience and clarity, we will discuss MCS and UWS as discrete entities, reflecting a presence or absence of conscious awareness respectively.

It is important to briefly describe locked in syndrome (LIS), which categorizes patients that are both aware and awake, yet have severely limited behavioral responses. Behaviorally unresponsive MCS patients are technically locked in, if they are fully aware, yet lack the ability to respond. Bruno and colleagues (2011) propose the term functionally locked-in syndrome for this patient population to differentiate them from classical locked in syndrome. Patients diagnosed with classical locked-in syndrome have a very limited behavioral response repertoire, usually in the form of eye movement. This syndrome is commonly caused by an infraction of the basis pontis following a stroke (Bruno et al., 2011; Patterson & Grabois, 1986). This type of injury leaves the vast majority of the brain intact. This is in stark contrast with the two main etiologies leading to a UWS diagnosis, which both cause widespread neuronal damage. While neuroimaging paradigms have proved fruitful in assessing awareness within a subgroup of behaviorally unresponsive MCS patients, it is important to stress that these patients are far from homogenous. Unlike classical locked-in patients they commonly have fluctuating levels of vigilance, sensory defects and varying levels of awareness. This patient heterogeneity is due to the wide variety of etiologies that can lead to a DoC.

In an attempt to circumvent these confounds novel neuroimaging diagnostic methods have been proposed. Functional neuroimaging allows the ability to observe in vivo patients brains hemodynamic activity (via functional magnetic resonance imaging [fMRI] or H215O-PET),

electrical activity (via electroencephalogram [EEG] or magnetoencephalography [MEG]) or metabolic activity (via 18F-fluorodeoxyglucose PET [FDG-PET] or magnetic resonance spectroscopy [MRS]) (Giacino et al., 2014). One of the first and most consistent neuroimaging finding in UWS patients is a global decrease of around 50% in overall brain metabolism, measured by FDG-PET (Laureys et al., 2002). The figure below (Fig. 2) illustrates how different neuroimaging modalities offering can offer a variety diagnostic information in DoC patients.

Figure 2 Reproduced from Bruno et al. (2011) under the creative commons license. This figure shows

the utility of multiple neuroimaging modalities (FDG-PET, default mode fMRI, high-density EEG & diffusion tensor imaging) in diagnosing patients with DoC.

Recently, researchers have developed paradigms to determine the presence of consciousness in DoC patients. These paradigms can be roughly broken up into two categories, reflecting their demands upon the patient, as active (command following) or passive paradigms (measuring the brains response to external sensory stimulation) (Heine et al., 2015; Kondziella, Friberg, Frokjaer, Fabricius, & Møller, 2015). We will expand the definition of passive paradigms to include any experimental design that does not explicitly instruct the participant to perform a certain task, this allows for the inclusion of resting state studies (Gosseries et al., 2014; Monti, Laureys, et al., 2010).

Owen and colleagues’ (2006) seminal study serves as an appropriate example of a passive and active fMRI paradigm for the assessment of consciousness in a UWS patient. The first task (passive) was used to measure the patient’s neural responses to spoken sentences while contrasting to matched noise sequences. Speech specific activity was similar to that of controls. Yet an important caveat was that some of these sentences contained intentionally ambiguous words, such as “The creak came from a beam in the ceiling” (Owen et al., 2006, p. 1402). These ambiguous words produced an additional significant response which was similar to healthy controls. This additional neural response, in the left inferior frontal region, was inferred as semantic interpretation. This highlights one of the critical limitations regarding passive

paradigms, which is a lack of confirmation of conscious awareness. Many studies have shown semantic processing can be present with the absence of consciousness, therefore the only way to accurately confirm awareness of a subject is through self-report (Kouider & Dehaene, 2007). The second paradigm (active) directly addresses this limitation by instructing the patient to either imagine playing tennis (motor imagery), imagine walking around her house (spatial imagery) or simply rest (baseline). The authors reported similar neural activity during visualization conditions as healthy controls [medial frontal regions for motor imagery and parahippocampal gyri for spatial imagery] (Owen et al., 2006). Owen et al. argue the patient’s ability to willfully modulate brain activity upon command represents a clear act of intention, therefore she must possess awareness. This is a clear example of reverse inference, since Owen and colleagues (2006) use the neural activation of a particular region to infer engagement of a cognitive task (Nachev & Husain, 2009). Their line of reasoning is inductive rather than deductive, yet since the results reflect the expected regions on the basis of prior literature this issue is slightly mitigated (Poldrack, 2006).

Before continuing, it is critical to introduce two central cognitive theories on consciousness, as these will help interpret later findings: global neuronal workspace theory (GNW) and integrated information theory (IIT). They are both models addressing how information gains access to conscious processing. Yet while both theories emphasize the integration of distributed information, they tackle question of consciousness from different directions. GNW theory starts from the brain and proceeds to explain how it can give rise to experience, in contrast IIT starts from the essential properties of experience (axioms) which leads to inferences about the characteristics of a physical substrate (Tononi, Boly, Massimini, & Koch, 2016). GNW theory posits that global workspace ignition is necessary for information to reach higher sensory areas and in turn gain conscious access (Dehaene & Changeux, 2011). Ignition is the process which makes information available to multiple brain areas, these areas then reverberate creating phenomenological experience. To clarify, reverberation is a term used for top-down feedback signals which form recurrent loops. When a stimulus does not reach conscious access it still can initiate feed-forward processing, in early sensory areas, without reaching higher associative areas and igniting the global workspace. Figure 3 illustrates the difference between feed-forward (unconscious) and an ignited reverberating global workspace. Dehaene and Changeux (2011) state consciousness is possible via a network of long range neurons, most heavily concentrated in the prefrontal, parietal-temporal and cingulate cortices.

Figure 3 Reproduced from Dehaene and Changeux (2011) under the creative commons license. This

figure illustrates unconscious feed-forward processing on the left and reverberating global workspace ignition on the right.

In contrast, integrated information theory (IIT) offers a mathematical framework to assess the quality and quantity of consciousness. The five axioms of IIT state every experience exists intrinsically and is structured, specific, unitary and definite (for a more detailed description on the axioms see Tononi et al., 2016). These postulates are universally applied to individual mechanisms and system mechanisms. IIT subscribes to physicalism, in that for each of these ‘essential properties’ of an experience there must be a causal physical substrate of consciousness (PSC), yet it does not necessarily have to have biological underpinnings. The parts within this PSC must also possess intrinsic cause-effect power, resulting in a cause-effect structure governed by the axioms (Tononi et al., 2016). This leads to the central tenant of the theory that an experience is a maximally irreducible conceptual structure (Oizumi, Albantakis, & Tononi, 2014). In a practical sense IIT dictates conscious processing to be differentially integrated (Boly, Massimini, & Tononi, 2009; Tononi, 2004). It is crucial to note that both theories stress the integration of distributed information as a prerequisite for conscious experience.

This manuscript will proceed by reviewing passive neuroimaging paradigms (Ch. 1). These studies dismissed the early idea of UWS patients lacking cerebral activity. Rather, passive paradigms revealed DoC patients to retain isolated higher level processes in ‘functional islands’. Yet for determining conscious awareness these paradigms lack confirmatory power. Therefore, chapter two will focus on active neuroimaging paradigms, with a particular emphasis on communication protocols. Studies that possess both an active and passive element will be discussed at the end of the chapter. Lastly, the manuscript will end by highlighting promising developments and future directions.

Passive Paradigms

To reiterate, we define a passive paradigm as any experimental design that the participant is not explicitly instructed to perform a certain task. Research on DoC patients has involved numerous types of passive paradigms using a variety of imaging modalities (Giacino et al., 2014; Harrison & Connolly, 2013; Kondziella et al., 2015; Monti, 2012). This review will not attempt to provide a comprehensive overview of each paradigm and modality, rather it will aim to provide a general summary of the research.

The early neuroimaging studies researchers designed to assess awareness in DoC patients utilized passive paradigms exclusively (Giacino et al., 2014). These studies immediately challenged the idea that UWS patients lack cerebral cortex activity. This view is clearly illustrated by one of the first names for UWS, apallic syndrome, derived from a-pallium meaning “without a cortex” (Monti, 2012). FDG-PET studies helped challenging this view by showing global cerebral cortex activity, albeit diminished (~50% when compared to healthy awake controls). Similar research also demonstrated that the fronto-parietal areas are the most hypo-metabolic in UWS patients (Laureys et al., 1999; Soddu et al., 2016).

This notion of an ‘apallic brain’, devoid of cerebral cortex function, was further challenged by H215O-PET studies. H215O-PET has the ability to measure local hemodynamic activity with

decent spatial and temporal resolution. In agreement with the earlier FDG-PET studies, it immediately became evident that not only there is cerebral cortex activity in UWS patients, this activity is also sensory specific (Monti, 2012). Jong and colleagues (1997) published one of the first studies to examine language processing with affective connotation in a UWS patient. In addition to sensory specific processing, the authors showed activation in areas known to be implicated interpretation and comprehension of affectively intonated speech. Quickly the extent of cerebral cortex activity became of interest: Are UWS patient’s neural activity confined to primary sensory areas or can information reach higher level associative areas? Over a decade of multimodal neuroimaging research has attempted to answer this question, using a wide variety of stimulus types (visual, auditory and tactile). Moreover, it has been assumed that if a stimulus can reach higher associative areas there is a distinct possibility that awareness is present in that patient. This assumption is in line with GNW theory, since the theory states that global workspace ignition (the causal process leading to consciousness) is necessary for information to reach higher sensory areas. Recently, researchers have proposed increasingly complex passive experimental designs to examine differing levels of semantic interpretation in DoC patients. An apt example of a multilayered passive paradigm is discussed in detail by Coleman and colleagues in 2007 and 2009. The paper published two years later adds more DoC patients from another location, yet the analysis and paradigm remain intact. These studies use the same passive paradigm from Owen and colleagues’ (2006) seminal ‘tennis imagery’ study, aforementioned in the introduction. The patient reported in Owen et al. (2006) study is included in the analysis of both proceeding studies. The overarching goal of all these studies was to assess language comprehension in DoC patients (UWS, MCS and

emerging-MCS). The experiment was designed in a graded fashion, in that each part became more semantically difficult as the patient proceeded. To reiterate, the first phase was a simple comparison between auditory stimuli and silence. This contrast allowed the authors to check for auditory sensory pathway integrity, which is critical in DoC patients since they commonly have sensory defects (Majerus et al., 2009). The middle phase contrasted the response of neural activity between intelligible speech and matched unintelligible noise stimuli. This contrast focused on the ability to differentially process speech. Lastly, higher level speech comprehension aspects were probed by contrasting sentences with semantically ambiguous words to those without semantic ambiguity (High ambiguity: e.g. ‘There were ‘‘dates’’ and ‘‘pears’’ in the fruit bowl’).

From a theoretical standpoint the emerging-MCS patients should show significant activation in all three levels of language processing, yet when corrected for multiple comparisons they did not show significant whole brain effects for the semantically ambiguous paradigm. The authors argue this result is due to the overly conservative nature of the analysis, therefore they proceed to reanalyze the data using a region of interest (ROI) approach. The regions used were chosen because they showed significant activation in healthy controls on the same task at the group level. While the authors state “this procedure did reveal significant increases in activity for semantically ambiguous sentences” (Coleman et al., 2007, p. 2499). Upon closer examination, of the emerging-MCS patients in the semantical ambiguous paradigm, only one (out of two) ROI in one of the patients (n=2) was significant. This analysis is quite important since it allows an objective measure of task validity in the highest semantic level of the paradigm. The lack of a strong result in emerging-MCS patients immediately calls into question their semantically highest paradigm (ergo the most important one towards assessing awareness) in subjects that are known to be aware. Three of the UWS patients (out of 7) and two of the MCS patients (out of 5) showed significant temporal lobe responses to the mid-level paradigm (speech vs noise). This result again calls into question the validity of the paradigm since three MCS patients failed to show significant differential activity between speech and noise, yet this result can be partially explained by fluctuating levels of arousal in MCS patients (Harrison & Connolly, 2013; Naci & Owen, 2013).

The most impressive result of the Coleman and colleagues 2007 paper is one UWS patients who “almost” showed whole-brain corrected significant activation in the ambiguous paradigm, yet this patient was previously reported (Owen et al., 2006). Overall the speech comprehension paradigm used by Coleman and colleagues does not provide a consistent and accurate method of assessing semantic ability in DoC patients. It is important to mention one of the main strengths of this paradigm, which is that it goes beyond simple group-based analysis and attempts to distinguish awareness on the individual level. Briefly mentioned in the introduction, one of the main limitations of the majority of passive neuroimaging studies is analysis at the group-level. If the results are analyzed at the individual level, any indication of awareness in the patient can provide additional (subject-specific) diagnostic information.

As briefly mentioned above, Coleman and colleagues (2007) choose regions of interest based on healthy controls at the group level, yet they do not report task performance of healthy controls at the individual level. Healthy controls and emerging-MCS patients are both suitable populations to test the validity of awareness assessing neuroimaging paradigms. Since both groups can reliably produce behavioral signs of awareness, one can use these populations to determine the amount of false-negatives (type 2 errors) in a particular paradigm. Emerging-MCS have further benefits since they possess similar structural damages and etiologies as UWS patients (i.e. they are a representative sample). Secondly, one of the most commonly cited explanations for false-negative findings of awareness in MCS patients, fluctuating levels of vigilance, is mitigated by using emerging-MCS patients.

Generally speaking, passive neuroimaging studies using semantic interpretation have been unsuccessful at producing accurate individual-based assessments (Gosseries et al., 2014; Kondziella et al., 2015). Naci and colleagues (2014; 2015) propose a novel passive method to assess awareness in DoC patients using captivating stimuli, such as watching scenes from a Hitchcock movie or listening to an excerpt from the movie Taken. One of the benefits of this type of paradigm is the engaging nature of the stimuli, as it helps capture attention naturally. This is particularly important for DoC patients as they commonly have limited attentional resources. Another, subtler benefit of this paradigm was the method of analysis, which focused on the temporal evolution of brain activity rather than activity in an expected region. Most neuroimaging paradigms to assess awareness in DoC focus on comparing regions with similar activation. Yet this approach is severely limited, since the precise spatial location of healthy subject’s neural activity deviates substantially at the individual level. Naci and colleagues (2015) state that within brain injured patients this issue is exacerbated due to substantial structural damage and the accompanying functional reorganization. Using this paradigm, the researchers could identify probable covert awareness in one UWS patient, leading them to claim that he was “continuously engaged in conscious thoughts about real world events unfolding over time” (Naci et al., 2015, p. 309).

Recent studies have made headway by using different techniques such as TMS-EEG and resting-state. TMS is a non-invasive task-independent tool for localized brain stimulation (using the principles of electromagnetic induction) that has a wide variety of uses in cognitive neuroscience (for a recent review on non-invasive brain stimulation see Parkin, Ekhtiari, & Walsh, 2015). Rosanova et al. (2012) published the first study attempting to longitudinally measure the recovery of effective connectivity in DoC patients with a passive TMS-EEG paradigm. The theoretical underpinnings of this approach have already been alluded to, specifically the necessity of multiple specialized modules of the thalamocortical system engaging simultaneously in casual interactions to create phenomenological awareness (Dehaene, Kerszberg, & Changeux, 1998; Seth, Izhikevich, Reeke, & Edelman, 2006). The essential characteristics of this mechanism are differentially integrated cognitive modules. Therefore, if the system is perturbed, in a conscious state, the response (measured via EEG)

should be disperse and sequential. The combination of TMS and EEG allows researchers to perturb and record the neuronal response in DoC patients.

In contrast, resting-state is a term used to describe paradigms in which the subject is allowed to simply rest and let the mind wander, which gives rise to spontaneous connectivity patterns (Soddu et al., 2016). Similar to the TMS-EEG paradigm, this ‘task-free’ neuroimaging method has clear advantages for DoC patients since it does not require co-operation and circumvents task related issues (such as sensory pathway integrity). A recent systematic review by Hannawi and colleagues (2015) on multimodal (fMRI, PET and SPECT) resting-state studies found decreased intrinsic functional and anatomical connectivity in the thalamus and PCC/precuneus complex of DoC patients (when compared to healthy controls). Yet resting-state studies possess the same inherent limitations as most passive paradigms studies, in that they both lack the ability to distinguish between MCS and UWS on the single subject level (Kondziella et al., 2015). The review concludes by emphasizing the need for more research, specifically on the topic of which functional activation patterns predict DoC recovery.

Another avenue for future research involves investigating the temporal dynamics of resting-state networks. The vast majority of resting-resting-state neuroimaging research, including the studies reviewed by Hannawi and colleagues (2015), derive conclusions of functional connectivity across the entire time course. Barttfeld and colleagues (2015) analyzed resting-state macaque data, under varying levels of awareness, to show two distinct origins of resting-state connectivity patterns: 1) continuous stream of ongoing cognitive processes and 2) random fluctuations shaped by a fixed anatomical connectivity matrix. The first origin, a stream of cognitive processes, should in principle have a clear time aspect, as the subject switches from one thought to the next. Barttfeld and colleagues (2015) showed these temporally limited ‘brain states’ to disappear under sedation. The authors identify the main implication for DoC research, which is the utility for this type of resting-state temporal analysis to identify ‘brain states’ in patients. If one could find distinct temporal connectivity patterns, which are only present in aware patients (MCS), the issue of incorrect diagnosis in DoC patients would be solved. In an attempt to develop a sensitive method, independent of the integrity of sensory or motor pathways, Rosanova and colleagues (2012) used TMS to perturb cortical neurons and record the response with high-density EEG electrodes. They found that in healthy awake subjects a TMS perturbation resulted in a sustained EEG response of disperse sequential brain activity, indicative of effective connectivity. Yet in slow wave sleep, an awareness absent state, this disperse sequential brain activity was lost (Massimini et al., 2005). The same research group, used the same procedure to examine effective connectivity in subjects following general anesthesia. The results were similar to slow wave sleep, which conceptually makes sense since they are both awareness absent states (Ferrarelli et al., 2010). Yet during rapid eye movement (REM) sleep, a state with phenomenological experiences (i.e. dreams), the TMS-triggered cortical activation was widespread and differentiated, similar to awake subjects (Massimini et al., 2010).

In 2013 the same group published a paper in which they derived an index for measuring cortical complexity (integration and differentiation), appropriately named the perturbational complexity index (PCI). PCI (range = 0-1) is defined as the “normalized Lempel-Ziv complexity of the spatiotemporal pattern of cortical activation triggered by a direct TMS perturbation” (Casali et al., 2013, p. 2). A high PCI results when the stimulation is propagated to a large set of integrated yet differentially reacting areas, which in turn gives rise to an irreducible spatiotemporal pattern. Casali and colleagues (2013) reanalyzed the previously mentioned results, in addition with novel data, to draw between-state conclusions using the PCI. Firstly, in healthy individuals the authors found significant differences in PCI values when contrasting wakefulness with NREM sleep, midazolam deep sedation or general anesthesia (propofol and xenon). As theoretically expected there was no significant difference of PCI between unconscious (i.e. awareness absent) conditions. In six subjects propofol was administered in a graded fashion, analysis determined significant effects of sedation levels on the PCI. Furthermore, the authors present even more evidence towards the ability of the PCI to distinguish between levels of consciousness by showing that it could differentiate a ‘transition state’ between wake and NREM sleep. As aforementioned, REM sleep showed wake-like effective connectivity, which was also reflected in the PCI.

The PCI was also able to discriminate between different disorders of consciousness at the group level, as seen in figure 4 (Napolitani et al., 2014). In UWS patients (n=6) PCI was within the same distribution as unconscious healthy subjects. Interestingly, two locked in patients (LIS) demonstrated a PCI equivalent to the value found in healthy awake subjects. Casali and colleagues (2013) also report that MCS patients had a significantly higher PCI than UWS patients. One of the most impressive findings is the PCI in emerging-MCS patients, which was significantly lower than LIS patients and higher than UWS patients. Yet when we consider the etiologies of these two patient groups, the result becomes less impressive. Since the brains of LIS patients are completely intact, it makes sense that recovering DoC patients with diffuse neuronal loss would have lower PCI’s. Therefore, it is important to note that a lower PCI score could reflect the neuronal injury profile of the patient rather than a difference of awareness (although in DoC these two are commonly correlated). The comparison between MCS and emerging-MCS did not produce significant results.

Figure 4 Reproduced from Napolitani and colleagues (2014) under the creative commons license. (A)

Displays locked in state, emerging-MCS, MCS and UWS patient’s averaged PCI scores per subject along with healthy subject distribution of wakeful and unaware PCI scores. (B) Shows the between-group significance. *P =0.002, **P = 0.0001

Overall the TMS-EEG technique (along with the PCI) seems to have more diagnostic utility than classic EEG techniques (Ragazzoni et al., 2013). Secondly, the site of TMS stimulation had no quantitative effect on the PCI value of a subject (Sitt, King, Naccache, & Dehaene, 2013). Further research is needed on larger sample sizes to determine the within/across subject reliability of the PCI for differential diagnosis’s in DoC patients. This is needed due to the low sample sizes (n=6) in each DoC group of Casali and colleagues (2013) paper. It is also important to mention the that PCI could potentially provide a quantitative measure for neurodegenerative diseases or symptomatic mild traumatic brain injury. The reason for this hypothesis is that cortical excitability has been shown to be altered in these diseases, therefore it may be possible to quantify disease progression with the PCI (Napolitani et al., 2014). Lastly, a recent paper by Guller and Giacino (2014) outlines the potential applications of a similar TMS-fMRI design. Substituting fMRI for EEG allows for a higher spatial resolution, yet this comes with a significant loss of temporal resolution. The temporal resolution (millisecond scale) of EEG is necessary to quantify causal interactions between different brain areas across time (used to derive a PCI), therefore the substitution of fMRI may diminish the benefit of this technique (Casali et al., 2013, Supplementary material).

From a diagnostic point of view, it does not matter if there is a significant DoC group difference using a particular neuroimaging method. A finding of this nature would not help in fixing the potential incorrect classification of MCS as UWS. For example, passive paradigms have consistently shown diminished neural activity in the fronto-partietal areas in DoC patients, yet this result is of little practical use (Bruno et al., 2010; Stender et al., 2014). What is currently missing and what would have a substantial impact is an accurate and robust behavior-independent measure with the power to discriminate awareness at the level of the individual (Harrison & Connolly, 2013). Briefly touched upon in the introduction, the presence of consciousness needs to be confirmed by the subject (e.g. “Yes, I am aware of my surroundings!”). Confirmation by the subject is the only way to prove that they had some sort of purposeful interaction with the environment or stimuli (Bodien & Giacino, 2016). Yet passive neuroimaging techniques possess the inherent flaw that they will never be able to provide certainty of the presence of awareness, because there is no behavioral confirmation by the subject. This unavoidable flaw in passive neuroimaging paradigms is circumvented with active neuroimaging paradigms. The next chapter will proceed to outline how active neuroimaging paradigms can allow a subset of behaviorally unresponsive patients to answer questions. The ability to answer questions in itself confirms the presence of awareness.

Active Paradigms

The birth of active neuroimaging paradigms to assess the presence of consciousness in UWS patients can be traced back to Owen and colleagues’ 2006 paper. In which an UWS patient was instructed to modulate her brain activity upon command to three conditions: visualizing walking around her house, visualizing playing tennis or resting. Using the neural activity from

healthy controls the authors hypothesized that visualizing navigating around a familiar environment would activate the parahippocampal gyrus, the posterior parietal cortex and the lateral premotor cortex. Alternatively, imagining playing tennis was hypothesized to activate the supplementary motor area. When the UWS patient was instructed to visualize motor activity or navigation the respective a priori regions showed significant activation. Owen and colleagues (2006) stated that ability to modulate brain activity upon command represents a clear act of intention and therefore the subject must be aware.

The results from active fMRI imagery paradigms have been extensively replicated in DoC patients (Bardin et al., 2011; Bekinschtein, Manes, Villarreal, Owen, & Della-Maggiore, 2011; Fernández-Espejo & Owen, 2013; Monti, Vanhaudenhuyse, et al., 2010; Noirhomme, Brecheisen, Lesenfants, Antonopoulos, & Laureys, 2015; Stender et al., 2014). Similar selective attention variants, in which the subject is instructed to modulate attention, have also proven fruitful (Monti, 2012; Naci & Owen, 2013; Schnakers et al., 2008). All command-based paradigms can be modified, through instructions, to reflect a binary choice, which in turn can facilitate communication. For example, the participant can be instructed to modulate neural activity using spatial navigation or motor imagery to reflect an answer (e.g. ‘visualize playing tennis to indicated a yes response’). Communication protocols start by asking factual questions about the patient to assess validity, an example can be seen in figure 5. Once the ability to communicate has been established, clinically relevant questions can be asked, such as “Are you in pain?” (Fernández-Espejo & Owen, 2013; Monti, Vanhaudenhuyse, et al., 2010).

Figure 5 Reproduced from Monti et al. (2010) under the creative commons license. A-D show differential

BOLD activation reflecting a response in a DoC patient and a control. Both participants were instructed to use motor imagery to indicate “Yes” and spatial imagery to indicate “No”. Widespread subcortical damage can be seen in the DoC patients T1 image.

Active paradigms are a promising development in the search for a robust diagnostic tool for DoC, yet two important questions arise: 1) What is the quality and extent of these patient’s experiences? and 2) Prevalence, how often do active paradigms identify a fully aware UWS patient?

Currently, it is impossible to gauge the extent or quality of these patient’s experiences, since communication is severely limited in a binary (yes/no) fashion. As for prevalence, a study conducted by Bekinschtein and colleagues in 2011 can offer a preliminary figure of 8.3%. The authors first used a passive EEG paradigm to assess auditory pathway integrity in patients, before conducting an active fMRI task. The active paradigm involved the patients imagining moving their right or left hand when instructed. In total 24 UWS patients were assessed, 13 of which did not proceed to the fMRI task due to a lack of EEG auditory response. Of the 11 subjects that were left, six were excluded due to either MRI artifacts or a lack of activation in auditory pathways. These exclusion criteria left only five subjects with auditory-specific

While larger studies have been conducted, the levels of occurrence reported are bias, since these populations were chosen by convenience at coma centers (Monti, Vanhaudenhuyse, et al., 2010; Stender et al., 2014). For example, Monti, Vanhaudenhuyse and colleagues report a slightly higher incidence rate of 5 aware UWS patients out of 54, yet this population was drawn from two major coma referral centers. One of the strengths of Bekinschtein and colleagues’ (2011) paper is that their patient population were consecutively admitted, allowing a more ecologically valid incidence rate (e.g. the subjects were not part of a subset of patients referred to a specialized coma center). A review by Kondziella and colleagues (2015) claims it is reasonable to assume that the true incidence of aware UWS patients, as determined by active

fMRI paradigms, to be around 5-15 percent. Interestingly, etiology is a major predictor of

which patients will show awareness. The vast majority of aware UWS patients, as assessed through neuroimaging, tend to have suffered traumatic brain injury (Monti, Vanhaudenhuyse, et al., 2010). This result is not very surprising since traumatic DoC patients have been shown to have a higher rate (12-month post injury) of recovery of consciousness (52% v 13%) in comparison to non-traumatic DoC patients (Monti, Laureys, et al., 2010). A possible reason for this effect is that TBI tends to inflict severe local damage rather than wide spread global damage.

fMRI is the most widely used neuroimaging modality for active paradigm research, yet this

does not necessarily mean it is the most ideal. In reality, the limitations that fMRI pose disproportionally affect DoC patients. Simple practicalities such as the necessity for dedicated infrastructure makes bedside testing an impossibility. In addition, fMRI is very susceptible to movement artifacts. This does not bode well for DoC patients as they commonly have issues with laying still for extended periods of time due to motor impairments (Bodien & Giacino, 2016). This issue is exacerbated in active fMRI paradigms, since sedation, commonly used to quell movement, is not an option as it would impair task performance. Other imaging artifacts, such as ones due to metal implants, are also commonly found in DoC patients. Due to these limitations only 30-60% of DoC patients can even be assessed with MRI (Noirhomme et al., 2015; Stender et al., 2014). Within this group of patients, limitations are still present, since the method of analysis may significantly differ depending on the group analyzing the data. fMRI has no standardized procedures for data preprocessing and analysis, therefore there is the risk of differing analysis procedures producing contradictory results. This lack of standardized procedures, effects the neuroimaging field as a whole and is partially responsible for the current reproducibility crisis (Nieuwenhuis, Forstmann, & Wagenmakers, 2011; Stelzer, Lohmann, Mueller, Buschmann, & Turner, 2014; Turner, 2016; Vul, Harris, Winkielman, & Pashler, 2009).

Regardless of these limitations, fMRI has been the imaging modality of choice for the majority of active-paradigm DoC research (Harrison & Connolly, 2013). EEG has the benefit of being inexpensive (in turn allowing repeated measures and longer time courses), compact and readily available in the majority of clinics (Sergent et al., 2016). Similar to fMRI paradigms, active EEG paradigms broadly fall into two categories: imagery or selective attention. Selective attention paradigms benefit from being less cognitively demanding than imagery based paradigms, which is especially important in DoC patients (Noirhomme et al., 2015). Examples

of EEG selective attention tasks are counting the number of times the patient hears a specific target sound, or their ability to pay attention to global violations of temporal regularities of tone sequences (Bekinschtein et al., 2009; Schnakers et al., 2008, 2009). Command following is assessed by the ability of a subject to modulate attention upon instruction (i.e. to move from an inattentive to an attentive state or vice versa). Goldfine and colleagues (2011) preformed a study in which they analyze (with univariate and multivariate approaches) EEG power spectra changes during instructed imagery conditions (swimming or navigation) to rest in DoC patients and healthy controls. Since signal change patterns were inconsistent between controls the authors redefined a positive response as a consistent signal change (across tasks) within each subject. Differences were identified using the Two Group Test and were only considered significant if it they were present across all frequencies, over a 2hz range, contiguously (for more information see Goldfine et al., 2011). Using this criteria one MCS patient and one LIS patient showed evidence of motor task performance. Interestingly, of the healthy controls only 75% produced EEG data significantly above chance (Kondziella et al., 2015). This problem of failing to find significant neural activity in subjects clearly demonstrating behavioral awareness, is present in all active paradigms across differing imaging modalities.

An exchange between Goldfine et al. and Cruse et al. in 2013 epitomizes analysis issues in active paradigm research. Using more stringent statistical methods Goldfine and colleagues (2013) reanalyzed DoC imagery task data provided by Cruse et al. (2011). Specifically, Goldfine and researcher’s (2013) main issue with Cruse and colleague’s analysis method is that they did not account for correlations within (i.e. trials) or between experimental blocks. In DoC patients it is particularly important to correct for correlations between nearby blocks and trials, since patients commonly have fluctuating arousal states or artifacts. When reanalyzed, the data showed no evidence for task performance in DoC patients. Cruse and colleagues (2013) counter this finding in a reply by stating that Goldfine’s analysis method only detects command-following in 40% of healthy participants. Furthermore, Cruse et al. (2013) claim that

fMRI command-following has already been shown for two of the three patients reported to

have EEG command following. Yet it is important to note that the citation for this claim is Owen and colleagues seminal 2006 ‘tennis’ study, which reports on a single DoC patient. While Goldfine and colleague’s (2013) reanalysis methods, using cross validation and multiple comparison correction, are more statistically sound, one must consider the lack of detected command following in healthy individuals. Overall, this exchange highlights how important the type of statistical analysis chosen is for determining the presence of awareness. Since there is no gold standard for assessing the presence of awareness, the question of which analysis method is more appropriate for UWS patients will never be definitively answered. It is a quite common occurance that active paradigms (fMRI and EEG) fail to find activation in patients clearly demonstrating conscious awareness. Therefore, when analyzing it is important that researcher’s strike a balance between methods that are disposed to false-positives and those that are biased towards false-negatives (Giacino et al., 2014).

As previously mentioned in chapter one, a good way to assess the validity of a DoC paradigm is to attempt the same experimental design on healthy (i.e. aware) individuals, this logic also

study (n=32) in which they examine the best statistical method for correct and accurate decoding of motor (tennis imagery) and spatial (navigation imagery) BOLD responses in aware subjects. The active paradigm utilized was two-part and a replication of Monti, Vanhaudenhuyse, and colleagues’ (2010) paradigm. The first part was used to assess the ability of command following. This localizer task consisted of 5 alternating (between spatial and motor visualization) blocks divided by 30 second rest periods. The second task aimed at facilitating communication. For the communication phase subjects were divided into two counterbalanced experimental groups. One group was told if they wanted to answer “yes” to a question they should visualize moving around their house while the other group was told if they wanted to answer yes to a question they should imagine playing tennis and vice versa for submitting a “no” response. Individual and group-based task activation in the supplementary motor area (motor imagery) and parahippocampal gyrus (spatial imagery) can be seen in figure 6. Interestingly, at the individual level, 25% of the subjects showed no activation, in one or both of the expected regions, during the localizing task. Yet it is crucial to note these results did not hinder communication, since within this ‘lacking activation group’ 87.5% of correct answers were detected using the best classification method. This result is of particular importance since all fMRI DoC communication studies (including Bekinschtein et al. discussed previously) rely first on a localizer task. The common, yet incorrect, assumption is that a patient without significant activation in the ROIs are unable to use neural activity for communication. The paper also highlights the need for repeating questions since incorrect answers may be attributed to an absence of statistical power due to intra-individual variability rather than a lack of awareness. Lastly, the authors conclude by emphasizing that active paradigms can only be used for assessing the presence of consciousness, not its absence.

Figure  6  Reproduced  from  Comte  et  al.  (2015)  under  the  creative  commons  license.  The  first  row  shows  

activations  (at  group  and  individual  level)  of  the  SMA  during  motor  imagery  while  the  second  row  shows   PPA   activation   during   spatial   imagery   tasks.   The   group   analysis   contrasts   motor   imagery   to   spatial   imagery  while  the  individual  analysis  contrasts  both  imagery  conditions  to  baseline  (rest).

In general, it seems that active paradigms tend to be overly conservative (bias towards false-negatives) when diagnosing the presence of awareness in DoC patients. This is reflected by the large numbers of MCS patients (and healthy controls) which do not produce significant results in active paradigms. As mentioned previously, aphasia and other sensory deficits are common in DoC patients (Majerus et al., 2009). Therefore when developing active neuroimaging paradigms it is important to avoid over-reliance on a single sensory modality (Bodien & Giacino, 2016). Another limitation of active paradigms are the cognitive demands placed upon the patient. Most active paradigms require the coordination of multiple cognitive systems (e.g auditory, working memory, language, mathematics) to complete the instructed task. Imagery tasks are particularly cognitively taxing as they require extended mental control from the subject (Harrison & Connolly, 2013). Heavy cognitive demands can easily over exert UWS patients, leading to an inflated false-negative rate.

Despite all of these limitations, active paradigms offer an advantage that passive paradigms lack, that is definitive proof of the presence of awareness. Partially due to the strengths and weaknesses inherent to both types of approaches, Stender and colleagues (2014) examined a large cohort of UWS (n=41), MCS (n=81) and locked-in (n=4) patients with passive (FDG-PET) and active (fMRI mental imagery) paradigms. The researchers’ goal was to compare the diagnostic precision of the two imaging methods to a current behavioral scale (CRS-R) and a 12-month outcome scale (Glasgow Outcome Scale-Extended). To clarify, when using FDG-PET an UWS diagnosis was impingent upon complete bilateral hypo-metabolism of the associative frontoparietal cortex (i.e. no voxels with preserved metabolism). Any sort of metabolic preservation within these areas would result in a MCS diagnosis. For the active fMRI paradigm (similar to Owen et al. 2006) any task related activation in one of the expected anatomical areas would lead to a MCS diagnosis. The authors hypothesized that a MCS diagnosis is predictive of 12-month post assessment behavioral signs of awareness whereas a UWS diagnosis is predictive of no recovery. Stender and colleagues (2014) use this assumption so they can calculate specificity and sensitivity rates of the two imaging modalities, a task which is widely considered impossible (Gosseries et al., 2014). Kondziella and researchers (2015) go as far as to state “In the absence of a gold standard for consciousness, precise estimates of the sensitivity and specificity of active and passive paradigms are futile.” (p. 489). Therefore, the author of this review has chosen to not report any sensitivity or specificity estimates.

Even when excluding sensitivity and specificity estimates, Stender and colleagues (2014) paper is important because it compares two widely used active and passive neuroimaging approaches in a large DoC patient population. The results showed that within UWS diagnosed patients (n=41), 13 had contradictory (awareness present) imaging findings in at least one imaging modality (12 patients with FDG-PET and 3 with fMRI). Of the UWS patients indicated to be aware, nine recovered consciousness within a year, three died and one remained unresponsive. After excluding these nine subjects, no UWS patients recovered awareness one-year post diagnosis, supporting Stender and colleagues (2014) hypothesis. Furthermore, the limitations

population could be assessed with fMRI (due to the various sources of artifacts aforementioned). In contrast, FDG-PET could assess 92% of the patients. The authors state that active fMRI imagery paradigms should only be used in a complimentary fashion, since FDG-PET proved to be a more reliable and robust outcome predictor. Leading the authors to recommend a joint CRS-R/FDG-PET method for optimal DoC diagnostic power. This review will conclude by discussing the most promising methods and how these methods can propel future research in DoC.

Discussion

The aim of this review was to compare and contrast promising neuroimaging awareness assessing methodologies in behaviorally unresponsive DoC patients. UWS patients will remain problematic to diagnose until our understanding of the neural mechanisms underlying consciousness is expanded. Further mechanistic understanding will hopefully lead to a ‘gold standard’ approach for assessing consciousness. Currently the best neuroimaging methods to assess awareness in DoC patients utilize either passive (Ch. 1) or active (Ch. 2) paradigms. It is important to reiterate a main message of the review, which is that active and passive paradigms have different inherent limitations. One of the main limitations of active paradigms is that they place heavy executive demands on the patient, therefore only a minority of DoC patients can be assessed with these techniques. In contrast, passive paradigms do not place any demands on the patient. Yet this methodological benefit also contributes to an inherent limitation on the strength of conclusions one can draw from a positive (awareness present) result. Passive paradigms do not allow the patient to explicitly confirm the presence of their awareness. A confirmatory report of awareness by the subject is a necessary prerequisite to achieve certainty, since awareness is an internally mediated state.

The naturalistic paradigm designed by Boly and colleagues (2014; 2015), involving suspenseful audio/visual excerpts from movies, presents many benefits towards assessing consciousness in DoC patients. The executive demands placed on the patient are minimized, since the task is naturally captivating and passive in nature. In addition, the method of analysis focusses on the temporal dimension rather than spatial activation, therefore the time course of areas showing activation is of interest rather than simply the specific localization of activation. This type of approach mitigates some of the reproducibility issues with fMRI analysis, touched upon in chapter two. All of these benefits are accomplished without sacrificing within-subject resolution. Boly and colleagues’ (2014; 2015) paradigm cannot provide clear-cut certainty of awareness, as granted by communication protocols, yet the paradigm allows a reasonable extent of inference on the presence of awareness by comparing the time courses of activation across subjects (Sinai et al., 2017). Overall this method utilizes the benefits of active paradigms without placing heavy executive demands on the patient.

Another important consideration, commonly neglected when developing a neuroimaging paradigm to assess awareness, is ease of implementation in the clinic. EEG offers many of the

benefits of PET and MRI, yet without the extensive cost and necessary infrastructure. Since the field lacks a ‘gold standard’ of assessing the presence of awareness in DoC patients, Sergent and colleagues (2016) developed a multidimensional cognitive evaluation of awareness in DoC patients using EEG. The theoretical underpinnings, proposed by Bayne, Hohwy and Owen (2016), is that consciousness is multidimensional, therefore complementary information can be gained by combining multiple passive and active paradigms. This theory is in stark contrast with the traditional view that there are differing levels of consciousness (for the traditional view see figure 1), instead multidimensionality allows the possibility for considerable variation within a single DoC category (figure 7).

Figure  7  Reproduced  from  Bayne  et  al.  (2016)  under  the  creative  commons  license.  B)  A  conceptual  

sketch  of  multidimensional  global  states  of  consciousness  and  their  possible  relation  to  different  DoC.   This   figure   illustrates   how   there   can   be   considerable   variability   of   multidimensional   states   across   different  DoC.  

The aim of Sergent and colleagues (2016) study was to probe multiple cognitive processes through a single protocol in an attempt to improve the evaluation of DoC patients. The novel aspect of the study was to combine multiple paradigms (Active and Passive), with varying levels of cognitive demand, in an attempt to improve sensitivity. As previous mentioned, neuroimaging paradigms routinely fail at detecting awareness in healthy controls, this finding highlights the poor sensitivity and conservative (prone to false-negative findings) nature of these paradigms. The benefit of a multidimensional paradigm is that it allows for repeated assessments of awareness using different techniques, inherently make the assessment more sensitive (see control subjects in figure 8). One of the major strengths of the study was the authors experimental design choices, aimed towards ‘real-world’ implementation in the clinic. For example, practicalities such as the length of the protocol or construction of a corresponding analysis pipeline were implemented. Overall this study presents a foundation for improving the

Figure  8  Reproduced  from  Sergent  et  al.  (2016)  under  the  creative  commons  license.  A)  Average  scores  

for  the  3  respective  groups  plotted  on  a  radar/spider  plot  for  each  dimension.  B)  Bayes  factor  in  favor   for   hypothesis   1   (MCS>VS)   over   the   null   hypothesis   (MCS=VS)   and   the   alternative   contrast.   C)   Spider/radar  charts  for  each  subject  in  the  8  cognitive  dimensions.  

Reproducibility and replicability of results is important for the neuroimaging field in general, yet it is particularly relevant when studying DoC, since a differential diagnosis (between UWS and MCS) can have a large impact on patient care. As touched upon in the previous chapter, specific statistical choices made when analyzing DoC neuroimaging data may have serious consequences upon the results of a study. These statistical choices (e.g. to include permutation or multiple comparison correction) drastically impact the sensitivity and specificity of the

paradigm. Therefore, it is important that researchers try to strike a balance between analysis methods prone to false-positives and those prone to false-negatives.

Novel multivariate neuroimaging analysis approaches can mitigate some of the issues associated with classical univariate statistics. Noirhomme and colleagues (2017) present a paper in which they review the application of machine learning techniques in DoC research. Machine learning approaches combine information in a multivariate way, in some cases this allows the detection of differences that would be missed when using univariate methods. The authors outline two types of approaches in DoC research, 1) to detect changes of activation during tasks in DoC patients and 2) to use data from multiple subjects to build a model for individual based prediction. The benefits of the first approach, detecting changes of task elicited activation, is that it is not based on a prior hypothesis of location or time. The lack of a prior hypothesis in analytic methods is quite advantages, especially with pathological patients due to the possibility of functional reorganization. Furthermore, it circumvents some of the analysis issues previously discussed. The second approach is to use data from multiple DoC patients to train a multivariate model. This model is then used, at the subject level, to predict the presence or absence of awareness (UWS vs MCS). One limitation of this approach is that the labels fed into the model could be incorrect, for example in the case of a behaviorally unresponsive MCS patient diagnostically labeled as UWS. The reliability issues of behavioral DoC diagnoses have been outlined extensively in this review and present an unavoidable limitation for classification models. Interestingly, this type of approach can utilize a wide variety of features. For example, differing neuroimaging modalities can be combined into a single predictive model (Noirhomme et al., 2015). Both types of approaches have shown promise for the advancement of DoC awareness assessing neuroimaging paradigms.

Combining different types of data into one predictive model has the benefit of capturing different aspects of consciousness and analyzing them jointly. It would be feasible to combine Sergent and colleagues (2016) multidimensional paradigm with EEG measures from resting state and TMS. This would result in an extensive protocol using ten distinctive EEG markers of awareness. Researchers could then build an analysis pipeline using a predictive multivariate model to maximize statistical power. Stender and colleagues paper (2014) used 12-month post assessment behavioral awareness as an indication of covert awareness. The general idea being if a UWS patient shows covert awareness they are likely to recover within a year. This hypothesis held true for their sample (n = 41) of UWS patients (i.e. all UWS patients determined to possess covert awareness showed behavioral responses 12-months post neuroimaging assessment). Therefore, a 12-month post assessment behavioral scale can help ground truth the multivariate model. After extensive model training the data of an individual patient could be sent to a remote server for automatic diagnosis. Furthermore, the data could be fed back into the model to enhance accuracy. It is important that any paradigm built to assess awareness in DoC patients has the ability for widespread clinical implementation. Realistically, this paradigm would need a specialized nurse for EEG and TMS placement and would take a minimum of two hours.

While this review has focused on neuroimaging paradigms to assess consciousness, it is important to take a step back and reflect on the reasoning behind such a task. The societal implications between a UWS and a MCS diagnosis are quite severe, as this boundary determines the legality of euthanasia in some countries. One of the broad messages this review attempted to convey is that DoC exist on a continuum, yet the laws in place are clear cut and have not been revised to incorporate current scientific knowledge (Syd & Johnson, 2016). It will never be possible to know if a neuroimaging paradigm can determine the true amount of behaviorally unresponsive aware patients, yet through evaluation we can determine which method is optimal. Hopefully a multidimensional approach, in tandem with machine learning techniques will be able to identify the majority of covertly aware DoC patients.