Novel Approaches to Prediction in Severe Brain Injury

Structured Abstract:

Purpose of review:

Recovery after severe brain injury is variable and challenging to accurately predict at the individual patient level. This review highlights new developments in clinical prognostication with a special focus on the prediction of consciousness and increasing reliance on methods from data science.

Recent findings:

Recent research has leveraged serum biomarkers, quantitative electroencephalography, magnetic resonance imaging physiological time-series to build models for recovery prediction. The analysis of high-resolution data and the integration of features from different modalities can be approached with efficient computational techniques.

Summary:

Advances in neurophysiology and neuroimaging, in combination with computational methods, represent a novel paradigm for prediction of consciousness and functional recovery after severe brain injury. Research is needed to produce reliable, patient-level predictions that could meaningfully impact clinical decision making.

Introduction

Despite advances in care, the burden of severe traumatic and nontraumatic brain injury (designated here collectively as ‘acute brain injury’ [ABI]) remains high both in terms of mortality and long-term disability. With advances in intensive care, withdrawal of life sustaining-therapies (WLST) is emerging as the most common proximate cause of death in comatose patients with acute brain injury [1]. This raises serious ethical concerns, since current approaches to predict recovery of consciousness and functional independence lack accuracy.

Existing models applied in severe TBI and post-cardiac arrest populations perform reasonably well in predicting outcomes in low and high severity patients, but they do not afford precise outcome prediction for patients in the intermediate severity stratum [2]. Models for post-cardiac arrest neurological prognostication have false positive rates as high as 15% in retrospective validation studies—well above the 1% tolerance for false positive WLST recommendations that was suggested in a recent international survey of clinicians[3]. Taken together, the accuracy of current approaches is insufficient to enable prediction at the individual patient level

This review will address challenges inherent in prognostication after ABI, as well as opportunities to enhance the effectiveness of modelling and outcome prediction. After discussing the choice of outcome measures and timing, the connection between recovery of consciousness and functional recovery will be assessed. Then, advances in predictive methods will be reviewed by category. Models integrating high-dimensional data in which the number of features far exceed the number of subjects or observations will be reviewed.

Assessment of consciousness

From the clinical standpoint, consciousness may be viewed as a continuum of behavioral phenotypes ranging from unarousable unresponsiveness (i.e. coma) to full behavioral expression of consciousness (i.e. normal arousal and awareness). There are two important intermediary states of awareness. The vegetative or unresponsive wakefulness state (VS/UWS) is characterized by wakefulness in the absence of reproducible and purposeful behavioral responses to external stimuli. Patients in the minimally conscious state (MCS) have definite—yet often subtle and inconsistent—behavioral evidence of self or environmental awareness. This state is important to identify, as it carries a more favorable prognosis than VS/UWS [4–6].

Discrimination of levels of consciousness at the bedside typically relies on a hierarchical characterization of responsiveness and command following. The Glasgow Coma Scale (GCS) assesses consciousness indirectly through eye opening, motor responses, and verbal responses to stimuli and may be expressed as a composite score or as individual subscores. It is primarily used in acute management and mortality prediction [7]. The Coma Recovery Scale-Revised (CRS-R) assess levels of auditory, visual, motor, and verbal function as well as level of arousal and communication during a standardized exam [8]. Recently, complementary behavioral responses such as suppression of the auditory startle reflex and differential sniff responses to olfactory stimuli have been shown to identify level of consciousness and predict 6-month recovery of consciousness in small patient cohorts [9*, 10**].

Conscious awareness can be detected in behaviorally unresponsive patients using advanced electroencephalography (EEG) or magnetic resonance imaging (MRI) techniques. Patients with cognitive motor dissociation (CMD) may appear clinically comatose wakeful yet unresponsive (VS/UWS), but nonetheless show evidence of consistent command following on non-invasive neuroimaging tests [11]. The prevalence of CMD is high; appearing in at least 15–25% of acute to subacute ABI patients [12–16**]. Importantly, unresponsive patients in the CMD state, when compared to patients in whom CMD is not detected, have a higher likelihood of functional recovery [13*, 15, 16**].

Patients with chronic disorders of consciousness (DoC) have variable levels of behavioral arousal throughout the day and assessment of awareness is further complicated by factors such as sedation, pain, and pharmacotherapy in hospitalized, acutely injured patients [13*]. Thus, repeated assessments are recommended to reduce the likelihood of misclassification (e.g. UWS vs MCS) [17, 18].

Functional outcome measures

Early studies of traumatic brain injury focused on mortality as a primary outcome. However, in more recent work there has been an emphasis on measures of recovery of motor function, conscious awareness, and independent function as outcomes for predictive models.

The Glasgow Outcome Scale (GOS; “1-dead” to “5-recovery”) is a 5-point scale of functional recovery used extensively in TBI research. Given the GOS’ limited ability to resolve meaningful outcome categories, the 8-point GOS-Extended (GOSE) was developed, which has better psychometrics and is validated for clinician and surrogate telephone assessment [19, 20]. The cerebral performance category (CPC; “1-good cerebral performance” to “5-brain death) is an analogous scale used post-cardiac arrest anoxic injury research. The CPC has been criticized for lack of formal administration instructions [21, 22]. While these measures are relatively easy to collect, their ability to accurately characterize functional outcomes is limited. Functional independence is a category in the GOS, GOS-E and CPC, but these scales are too coarse to capture specific features of the functional state. For example, an individual with CPC 3 outcome may be either minimally conscious and completely bedbound or ambulatory and fully conscious but unable to safely care for themselves given difficulties with impulsivity or attention.

To further characterize functional status, several standardized measures have been developed. The Barthel Index categorizes patients according their ability to accomplish activities of daily living (ADLs) necessary for safe self-care (e.g. feeding, bathing, dressing, mobility, and continence/toileting) (“0-totally dependent” to “100-totally independent” [23]. Alternatively, the modified Rankin scale (mRS; “0-no symptoms” to “6-dead”) is used heavily in stroke recovery and clinical decision making, more crudely assessing overall physical and psychosocial dependence in daily living [24]. The Disability Rating Scale (DRS; “0-no disability” to “29-vegetative state”) describes the impact on instrumental activities of daily living in the chronic recovery state [25].

Of note, these scales are scored on a non-linear set of ordinal integer values. Frequently, investigators will collapse scores into dichotomized outcome measures: “good” outcomes with an element of functional independence (ex. CPC 1–2, mRS 1–3, GOS-E 4–8) and “poor” outcomes including death, unresponsiveness, complete dependence in activities of daily living (ex. CPC 3–5, mRS 4–6, GOS-E 1–3). While such binary outcomes can be incorporated in logistical models, non-parametric ordinal analyses may be more statistically efficient [26]. Most importantly, since acceptable standards of living vary widely, models trained on binary outcomes often cannot inform meaningful distinctions for families and surrogates who may have different perceptions of what constitutes a “good” outcome.

Cognitive outcomes

Patients who regain consciousness after ABI may experience varying levels of impairment in multiple cognitive domains including attention, memory, and executive function, and may have symptoms characteristic of mood disorders [27]. Their residual cognitive recovery can be remarkable; 22% of patients with severe TBI discharged unresponsive to stimuli eventually returned to school or work [18] and over a quarter of patients admitted to rehabilitation after moderate to severe TBI had disability free (DRS 0) recovery, with higher pre-injury educational status predicting better recovery [28]. There is a need for models to select ABI patients who are most likely to benefit from intensive neurocognitive rehabilitation and other therapeutic interventions in the acute and post-acute phase.

Predicting response to targeted therapies

Despite considerable preclinical and clinical research, therapeutic interventions known to significantly modify outcomes after ABI are limited in number. Evidence suggests that neurostimulants, and both invasive and non-invasive neuromodulation could be effective in certain populations. Biomarker-driven predictive tools are needed to select patients who are most likely to benefit from specific neurostimulant or neuromodulation therapy[29]. For example, quantitative electroencephalography (EEG) patterns may predict recovery after administration of amantadine [30] or correlate with behavioral improvements attributed to zolpidem [31]. Similarly, metabolic patterns may also predict response to L-DOPA in patients with MCS after traumatic brain injury [32, 33], while measures of structural axonal integrity correlate with response to non-invasive transcranial stimulation therapy after severe TBI [34].

Timescale of recovery

The timescale of recovery of consciousness after ABI is variable and a recent body of evidence indicates that meaningful improvement may be observed far beyond previously conceived time windows of recovery). In addition to early recovery days to weeks in ABI, recovery of functional communication and activities of daily living has been observed from months to a decade after post-arrest anoxic injury [35], TBI [36, 37], and across other etiologies of ABI [18, 38]. These data have prompted the American Academy of Neurology to discourage use of permanent vegetative state in favor of chronic VS/UWS [18].

Novel techniques and modalities for prediction

Markers of organ function are increasingly used in the prediction of outcomes in medical illness, but no serum or cerebrospinal fluid biomarkers has been shown to reliably predict recovery of consciousness, cognition, or functional independence in DoC patients after ABI.

Serum biomarkers

Previously studied markers in TBI include compounds associated with neuronal cell damage (e.g. neuron-specific enolase) and glial damage (e.g. glial fibrillary acidic protein) [39, 40]. Studies in patients with anoxic and traumatic injury suggest that neurofilament light protein outperforms CT scan and other biomarkers in predicting clinical outcomes [41**].

Biomarkers have also been explored to evaluate the degree of post-injury inflammation. Inflammatory markers such as interleukin-6 and tumor necrosis factor-alpha are elevated in the acute setting of traumatic brain injury. They can remain elevated over decades and may be associated with lasting cognitive impairment [42]. Other markers such as procalcitonin have weak or conflicting reports of prognostic value [43, 44].

Levels of tau protein correlate with poor long-term neurocognitive outcome after severe TBI [45] and cardiac arrest [46], suggesting that it indexes secondary neurodegeneration. Levels of tau are associated with worse structural damage and neurocognitive performance in chronic TBI, and tau deposits are colocalized with white matter destruction in patients who develop severe posttraumatic neuropsychiatric disease [45].

Electrophysiology

Clinical electroencephalography (EEG) is generally used to diagnose seizures and monitor sedation, but it may also be leveraged to predict coma outcome. In acute post-cardiac arrest coma, suppression of >50% of the EEG record predicted poor outcome of CPC 3–5 at 6-months, while continuous patterns at 12 hours after arrest were specific for a “good” outcome of CPC 1–2 [47*]. In the chronic stage of DoC after ABI, clinical grading scores correlate with command following and predict recovery on CRS-R scale on the time scale of months to years [48–50].

Bilateral absence of the somatosensory evoked response N20 component is one of the strongest predictors of poor functional recovery following cardiac arrest [51]. Cognitive auditory evoked responses may offer added power to predict emergence from MCS years after ABI [52].

Quantitative analysis of EEG data offers additional information for prognostic modeling. For example, EEG features of cortical connectivity readily distinguish among levels of consciousness and are reliable over repeated measurements [53]. Response to commands on EEG (or functional MRI) has been used to distinguish CMD patients from unconscious patients and are shown to predict long term functional outcome [13*]. New methods to assess language processing passively by the interaction of speech amplitude and EEG signals may have diagnostic potential [54, 55] and may also predict conscious recovery [54].

Neuroimaging

Computed tomography (CT) and magnetic resolution imaging (MRI) are used to inform prognosis in anoxic post-cardiac arrest coma. Cytotoxic edema in the first 72 hours after arrest is reflected in gray-white matter ratio on CT and diffusion restriction on MRI. Both of these imaging modalities can predict functional outcomes at a population level [56–60]. Quantitative analysis of anatomically segmented head CTs was recently demonstrated as a viable initial approach for early neurological prognostication following cardiac arrest [56].

Diffusion tensor magnetic resonance imaging, which quantifies white matter tract injury, predicts poor 6-month post-cardiac arrest neurological outcome (defined as CPC 3–5) with high accuracy [61]. In TBI, however, heterogenous patterns of MRI-defined injury may limit meaningful prognostication at the individual patient level despite standardized radiographic grading protocols [62].

Complementing these structural imaging modalities, functional MRI (fMRI) maps regional cortical activity from stimulus or task-induced variations in blood oxygen delivery (blood-oxygen level dependent signal; BOLD). Analysis of BOLD responses to command following and motor imagery tasks can identify CMD patients in the acute setting [12]. Another fMRI technique, more broadly applicable to the DoC population, is to analyze the spatial distribution of BOLD signal covariance in the absence of any structured task, the so-called “resting state” paradigm. This method has demonstrated discrete large-scale functional networks which have emerged as valuable markers in patients with DoC. As an example, recovery of “default mode network” functional connectivity appears to correlate with re-emergence of consciousness in severe TBI patients [63]. fMRI connectivity patterns can predict 3-month GOS scores in heterogeneous ABI populations [64] and 12-month CPC scores in post-cardiac arrest patients [65, 66].

Multimodal physiological measures

Invasive measurements of intracranial pressure, interstitial metabolites like glucose, lactate, pyruvate, glycerol, and glutamate, as well as measures of oxygenation, blood flow, and intracranial electrical activity are employed in some intensive care units to guide management of ABI at the bedside. While group-level associations of specific measures with poor neurological outcomes and mortality have been reported [67], these technologies seem most valuable to guide targeted therapies for short-term optimization of brain physiological parameters (ICP, brain interstitial lactate/pyruvate ratios) potentially affecting long term outcomes [68].

Big data approaches to predictive models

The vast quantities of data generated by technological advances in neurocritical care pose a significant challenge for conventional statistical methods. The development of artificial intelligence and machine learning specifically represent an opportunity to efficiently process large and heterogeneous data, and to improve prediction of recovery of consciousness or functional outcomes in the future [69].

Dimensionality reduction

When analyzing complex data, it is advantageous to reduce the dimensionality through techniques like independent or principal component analysis. This is particularly true when searching for candidate biomarkers in large collections of biological samples [70–72]. Dimensionality reduction may also support understanding of a biomarker’s role in the pathophysiology of brain injury. For example, neurofilament light chain values appear distinct from a handful of other markers, suggesting a secondary or reactive role after injury [39, 73].

Dimensionality reduction is a specific type of unsupervised learning, which can be used to reveal “hidden” patterns in data not readily apparent from descriptive statistics. For example, an automated clustering algorithm detected five separate phenotypes of anoxic brain injury after cardiac arrest, one cluster showing possible benefit of early coronary revascularization [74].

Advantages of automated machine learning

Machine learning approaches to analyze high-dimensional data may match or exceed the prognostic accuracy of expert clinicians. For example, automated models to predict functional outcome from clinical EEG baseline features [47*] and stimulus reactivity [75] outperformed expert review. ML-derived EEG features were similar to traditional clinical EEG features, suggesting that these models converged on similar measures as neurophysiologists [76].

Machine learning techniques can generate superior prediction when compared to traditional multivariable logistic regression models, as seen in comparison of a single-center severe TBI cohort [77]. However, this advantage disappears when the models have low number of variables, suggesting limited utility outside of high-dimensional datasets [78, 79].

Machine learning may be of particular value in predicting clinical outcomes from heterogenous data such as structural MRI. One ML model was able to differentiate lesions likely to cause deficits in either working memory or reasoning, suggesting that ML-driven structural imaging analysis may predict individual profiles of cognitive disability in addition to motor impairment [80].

Frequent sampling of high-dimensional data

Big data analysis techniques are well-suited for temporally dense sampling of observations such as neurophysiological time series, thereby increasing the chances of capturing the complex dynamics of neural recovery and fluctuating levels of arousal. In one report, machine learning analysis of multiple serum neuron-specific enolase levels after injury led to more accurate outcome prediction than conventional analytical methods [81]. This was also observed in machine learning models integrating physiological time series data, which had improved prognostication accuracy for post-cardiac arrest anoxic injury when compared to models that did not integrate physiological time series [82].

Conclusions & future directions

The goals of prediction in severe brain injury are to produce accurate and reliable, patient-specific prognostication of recovery of conscious awareness and functional independence. Among patients who have regained consciousness, there is a need for discriminative models to predict domain-specific neurocognitive function. Identification of conscious awareness and cognitive-motor dissociation remains a critical task, as these patients may benefit from early access to neurorehabilitation and the use of brain-computer interface technologies. Last, prediction models are also needed to help in the selection of patients most likely to benefit from therapies.

Current models to predict outcome of patients with ABI have limited accuracy due to infrequent assessments of conscious awareness, misdiagnosis of CMD patients, overly simplistic binary clinical outcomes at time points that may precede meaningful recovery, and inefficient statistical models. Improved prognostic performance might be achieved by integrating rich datasets including granular phenotype measures, behavioral assessments, EEG, and neuroimaging, incorporating data across many time points, and by appropriately trained and validated machine learning algorithms.

Key points:

• No prognostic tools have been validated for prediction of individual patient outcomes after severe acquired brain injury

• Existing prognostic studies have important limitations in the number of observations and validity of functional outcome scores

• Patients experience fluctuating levels of consciousness after brain injury, and may harbor covert conscious awareness not detectable on behavioral assessment

• New computational techniques show early promise for the development of prognostic models